The study is a sampled survey rather than a census of every unrestrained dog in San Antonio, and the announced project is a test of the surveying technology rather than a report of completed findings. The automated effort will revisit the 32 study locations used in a manual survey conducted in 2019. In that earlier survey, volunteers helped San Antonio Animal Care Services identify unrestrained dogs and classify them as “stray” or “roaming.” Under the survey’s stated criteria, strays appear malnourished and skittish, whereas dogs that seem healthier and more approachable are likely to be owned by local families and are classified as roaming.

The new study uses the same general sampling areas to determine whether a camera-and-software workflow can help characterize this population while retaining human review. For each deployment, SwRI engineers will use a cellular phone mounted to a vehicle window to gather video of unrestrained dogs. Object-detection software will identify candidate animals in the footage, after which automated and human review will examine the clips.

Reviewers are expected to verify that detected animals were unrestrained, determine whether a human was present, count the dogs and assess whether each animal was owned or a stray. The survey has an explicitly limited geographic scope. The 32 locations represent areas in which unrestrained dogs are common, but the sample does not cover the entire city.

The project will therefore produce information about dogs observed in the selected areas rather than a precise total for San Antonio as a whole. Its workflow will also connect machine vision with human evaluation: the phone-mounted camera gathers roadside video, the software detects animals, and reviewers determine whether each detection meets the survey’s criteria. No totals, detection rates, classification accuracy or comparisons with the 2019 manual survey are available at the start of the announced October-to-December 2026 research period.

Those outcomes have not yet been reported, and the release does not provide performance measurements for the system. San Antonio Animal Care Services Director Jon Gary said the study would give the department another tool for understanding where unrestrained dogs are seen while supporting its broader work to protect the health and safety of people and pets. Mike Koets, a staff engineer in SwRI’s Intelligent Systems Division, said the organization was excited to test the technology and help San Antonio apply real-world solutions to an issue affecting many residents.

The planned final report is expected to include summary data and videos of dogs observed during the deployment. Until that report is completed, the release supports a description of the survey method and its intended municipal application, but not conclusions about how many unrestrained dogs are present or how well the automated system performs compared with the earlier volunteer-led survey.

Subject of Research: A machine vision and human-review survey of unrestrained dogs in 32 sampled areas of San Antonio

Article Title: SwRI applies automation to count unrestrained dogs in San Antonio

Article References: SwRI applies automation to count unrestrained dogs in San Antonio. Original publication (Bibliographic details unavailable.)

Image Credits: AI Generated

DOI: Not provided

Keywords: unrestrained dogs, machine vision, San Antonio, Southwest Research Institute, object detection, stray dogs, roaming dogs, municipal surveys, animal care services, SwRI, applies, automation

Synthetic aperture radar can produce two-dimensional observations regardless of daylight or weather, making it useful for maritime monitoring, but ship targets are difficult to distinguish from the surrounding sea and coast. Their radar signatures can appear as scattered bright points rather than continuous objects, while speckle noise blurs boundaries and distorts local texture. Detection is further complicated by an exceptionally wide target-size range and by settings ranging from isolated fishing boats in open water to crowded berths, islands, buildings and other coastal clutter. The study addressed these problems by modifying three parts of YOLOv8: how features are extracted, how useful features are selected during multi-scale fusion and how predicted boxes are matched to ship shapes.

Coordinate attention, or CA, was added throughout the network’s neck, where features from different spatial scales are combined. Unlike channel-only attention, which summarizes a feature map without retaining its position, CA pools information separately along the horizontal and vertical axes. It then applies learned weights that preserve coordinate cues while identifying relationships among feature channels. In the model, this mechanism was intended to strengthen valid ship responses and suppress sea-surface speckle, nearshore structures and other distracting patterns. Ablation tests showed why placement mattered. Adding CA to the backbone produced precision of 0.878 and recall of 0.874, both below the baseline, while placing it in the neck yielded 0.891 precision and 0.923 mAP@0.5. The neck-embedded version also improved mAP@0.5:0.95 from 0.525 to 0.531 without changing the reported 909-frame-per-second throughput.

The second modification replaced one standard convolution in the C2f feature blocks at the P4 and P5 backbone stages with dynamic snake convolution, or DSConv. Standard convolution samples a fixed rectangular grid, which can poorly follow the curved or irregular outlines of vessels in radar imagery. DSConv first learns offsets that describe changing contour trends, then moves its sampling positions toward informative target regions. The researchers argued that this flexible pattern can preserve fine scattering details more effectively than fixed sampling while avoiding some background responses. Comparisons with four alternative convolution modules supported that interpretation. DSConv produced 0.901 precision, 0.899 recall and 0.929 mAP@0.5, outperforming the baseline and the SPDConv, WTConv and SAConv alternatives on conventional detection accuracy. Its strict localization score of 0.495 was nevertheless below the baseline’s 0.525, and its reported throughput was 476 frames per second.

The third component, Shape-IoU, replaced the baseline’s complete intersection-over-union branch for bounding-box regression. Shape-IoU combines overlap, scale deviation, center distance and explicit aspect-ratio consistency, with weights derived from the width and height of each ground-truth box. The intended effect is to favor predicted boxes that resemble the geometry of actual ships rather than rewarding overlap alone. Used independently, the loss raised precision to 0.895, recall to 0.897, mAP@0.5 to 0.919 and mAP@0.5:0.95 to 0.531. The authors noted, however, that ships with blurred or low-contrast boundaries can have ambiguous physical outlines. Strong shape constraints may improve detection at a conventional IoU threshold while causing small boundary offsets that are penalized by the stricter 0.5-to-0.95 metric.

Joint ablation experiments exposed an interaction among the three optimizations that was not evident from their individual results. Coordinate attention alone, DSConv alone and Shape-IoU alone each maintained or improved strict localization performance. Combining CA and DSConv produced the study’s highest mAP@0.5:0.95, 0.532, but adding Shape-IoU reduced that value to 0.514. The complete configuration nevertheless produced the best precision, recall and mAP@0.5, reaching 0.908, 0.904 and 0.925. The authors describe this as a multi-module coupling effect and suggest that dynamic sampling, feature re-weighting and shape regularization may accumulate small boundary perturbations for speckle-blurred targets. Those perturbations have limited consequences at IoU 0.5 but carry greater weight when evaluation demands close pixel-level alignment across a range of increasingly strict thresholds.

The model was trained and tested on 39,729 chips generated from 210 SAR scenes in the updated 2021 SAR-SHIP dataset. The source scenes comprised 72 images from China’s Gaofen-3 satellite and 138 from the European Space Agency’s Sentinel-1 satellite. Spatial resolutions ranged from 1.7 to 25 metres, and the collection covered HH, HV, VH and VV polarizations in port, nearshore, island and open-sea environments. After cropping and deduplication, the chips were randomly divided at the image-chip level into training, validation and test sets containing 27,902, 5,980 and 5,980 images, respectively. Researchers removed geographic overlap and duplicate crops to reduce leakage, manually corrected ambiguous annotations and resized all inputs to 320 by 320 pixels. This dataset design increases geographic diversity, although chip-level random partitioning cannot represent every possible form of overlap between scenes.

Training used an NVIDIA RTX 4090 with 24 gigabytes of memory, PyTorch 2.5.0 and CUDA 12.1. All models ran for 200 epochs with a batch size of 16 and an initial learning rate of 0.0015. Precision and recall were calculated at a confidence threshold of 0.25, while mAP values were aggregated across confidence thresholds under the COCO evaluation protocol. The researchers compared the proposed detector with YOLOv5, YOLOv8, YOLOv10, YOLOv11, YOLOv12 and RT-DETR-L. At conventional thresholds, the improved model delivered the strongest reported results: its 0.925 mAP@0.5 exceeded YOLOv12’s 0.923, YOLOv10’s and YOLOv5’s 0.921, and YOLOv11’s 0.918. It also outperformed RT-DETR-L in precision, recall and mAP@0.5, while using 11.19 million parameters rather than the transformer model’s 31.99 million.

The computational cost remained close to that of the baseline: the improved model contained 11.19 million parameters and required 28.6 billion floating-point operations, compared with 11.126 million and 28.4 billion for YOLOv8. Runtime, however, was less favorable because offset prediction, dynamic sampling and feature interpolation add memory-access-heavy operations not fully represented by the theoretical FLOP count. Such irregular operations also benefit less from the highly optimized routines used for regular GPU convolution. The reported 588 frames per second therefore reflected only forward-pass inference, excluding image loading, preprocessing and postprocessing, and the authors treated throughput as a secondary indicator sensitive to hardware and system load. The results support the model for accuracy-oriented applications with moderate real-time requirements, not for systems requiring the baseline’s maximum processing speed or the precise localization captured by mAP@0.5:0.95.

Cross-dataset tests indicated that the improvements were not confined to the training source. Weights trained on SAR-SHIP 2021 were applied directly, without reported retraining, to the HRSID and SSDD datasets at the same 320-pixel input size. On HRSID, the improved model raised precision from 0.824 to 0.832, recall from 0.507 to 0.508 and mAP@0.5 from 0.544 to 0.547, while mAP@0.5:0.95 fell from 0.213 to 0.204. On SSDD, precision increased from 0.816 to 0.825, recall from 0.634 to 0.646 and mAP@0.5 from 0.684 to 0.692, with strict localization declining from 0.263 to 0.247. Visual examples across four polarization modes also showed more confident detections and fewer clutter-induced false alarms than the baseline. These findings support favorable transfer across differing SAR data distributions, but the experiment did not establish performance for every radar sensor, sea state or operating environment.

Subject of Research: Multi-module YOLOv8 optimization for ship detection in high-resolution synthetic aperture radar imagery

Article Title: Multi-scale ship detection with high-resolution SAR images based on improved YOLOv8

Article References: Liu, D., Chen, H., Chen, J., Ding, Y., Zhu, J., & Lei, P. (2026). Multi-scale ship detection with high-resolution SAR images based on improved YOLOv8. Frontiers in Marine Science, 13, Article 1943050. https://doi.org/10.3389/fmars.2026.1943050

Image Credits: AI Generated

DOI: 10.3389/fmars.2026.1943050

Keywords: synthetic aperture radar, ship detection, YOLOv8, coordinate attention, dynamic snake convolution, Shape-IoU, remote sensing, computer vision, object detection, cross-dataset generalization, Multi-scale, ship

Potato foliar diseases pose a threat to crop growth and yield, making timely and accurate disease detection important for early intervention and field management. Reliable detection remains challenging in field environments, where lesions may have varied edges, textures and irregular boundaries. FEPL-DETR is an adaptation of the efficient detection transformer RT-DETR, with three principal components intended to improve lesion representation, feature integration and training supervision.

Its first component is a Faster Frequency-domain Dynamic Convolution CSPDarknet backbone, which was built to adaptively represent lesion edges, textures and irregular boundaries. The second is BF-RepELAN, a feature network that incorporates RepNCSPELAN4 into a bidirectional multi-scale fusion topology. Together, these modules are intended to strengthen the representation of disease-related features at different scales.

The third component is DSI-GAM, a semantic-distillation framework adapted to FEPL-DETR. The researchers used it to strengthen semantic supervision during training. The model therefore combines a more adaptive visual backbone, multiscale feature processing and a training framework designed to provide additional semantic guidance.

On the DLOP dataset, FEPL-DETR reached 90.6% mAP50 and 65.4% mAP50:95. Compared with RT-DETR-R18, the model improved mAP50 by 5.7 percentage points and mAP50:95 by 0.6 percentage points. Its parameter count was 67.0% lower and its computational cost was 72.6% lower than that comparator, indicating a substantial reduction in model size and processing demand while maintaining the reported detection results.

The edge-device test found that FEPL-DETR processed 13.1 frames per second when deployed on an NVIDIA Jetson Orin Nano. This result demonstrates operation on a resource-constrained device in the study’s tested deployment setup. The reported speed should be understood in that context rather than as a general guarantee of performance across all hardware or field conditions.

External-dataset evaluations also indicated that the model could generalize beyond the DLOP dataset. The supplied research evidence does not specify the additional datasets, image sources or measured results from those evaluations, so the extent of that generalization cannot be independently characterized from the abstract. It does establish that the authors assessed performance beyond the primary dataset.

The work is a computational detection study, not a trial of disease-control outcomes. The reported results concern model accuracy, computational efficiency and inference speed; they do not establish effects on treatment decisions, crop yield or farm productivity. The article was published as open-access, peer-reviewed research in BMC Plant Biology and was available as an early version subject to further editing before replacement by the final Version of Record. Within the reported scope, FEPL-DETR offers a compact computer-vision approach for potato leaf-disease detection, including operation on edge hardware.

Subject of Research: Lightweight transformer-based detection of potato leaf diseases under complex field conditions

Article Title: FEPL-DETR: an efficient lightweight detector for potato leaf diseases under complex field conditions

Article References: Hou, Q., Zhang, Z., Tian, Y., Fu, Q., & Deng, L. (2026). FEPL-DETR: an efficient lightweight detector for potato leaf diseases under complex field conditions. BMC Plant Biology. https://doi.org/10.1186/s12870-026-10190-9

Image Credits: AI Generated

DOI: 10.1186/s12870-026-10190-9

Keywords: potato diseases, machine learning, object detection, FEPL-DETR, RT-DETR, edge computing, plant pathology, computer vision, crop monitoring, efficient detection, efficient, lightweight

The researchers trained and evaluated the models under a shared COCO-style protocol, using 896-pixel inputs and a confidence threshold of 0.25 for the paired comparisons. Their dataset divided into 1,772 training images, 694 validation images and only 45 test images using a track-id-majority policy intended to reduce identity overlap. Of all annotated ant boxes, 14.3% represented trophallaxis and 85.7% represented normal workers, making the biologically important interaction state a minority class. RF-DETR produced 412 true-positive, 77 false-positive and 75 false-negative detections on the test split, while YOLO26n produced 379, 115 and 108, respectively, at an intersection-over-union threshold of 0.5.

The largest improvement appeared in the recognition of rare trophallaxis interactions. On the track-majority test set, average precision for the trophallaxis class rose from 0.245 for YOLO26n to 0.361 for RF-DETR Small, a gain of 0.116. Normal-state average precision also increased. This pattern suggests that RF-DETR was especially sensitive to the close-contact cues needed to distinguish interacting workers, although the experiment cannot establish that any one architectural feature caused the improvement because the detector families used different training schedules, optimization procedures and confidence calibration.

That extra accuracy came with a substantial runtime cost. On the same hardware class, an Nvidia GeForce RTX 4070, YOLO26n ran at about 49 frames per second, while RF-DETR Small processed roughly 18 frames per second. The observed trade-off means YOLO26n was approximately 2.1 times faster in the reported inference benchmark. A complementary Faster R-CNN ResNet-50-FPN model produced intermediate detection quality—better than YOLO26n but below RF-DETR—while recording the lowest throughput, at about 9.5 to 9.6 frames per second. A larger YOLO26s variant also failed to improve on YOLO26n at the chosen operating point.

The study was designed around a reproducible resource rather than a new detection architecture. The authors released the annotated dataset, split manifests, evaluation scripts, trained checkpoints and an interactive visualization tool so that future systems could be assessed using the same definitions and data. Annotators labeled each ant with a tight bounding box and assigned one of two states: normal or trophallaxis. When oral contact was clear and biologically plausible, both workers involved in an exchange were marked as trophallaxis, while ambiguous cases were left in the normal class. Track metadata and TransT-assisted tracking in CVAT helped maintain identity consistency during annotation in crowded scenes.

The recordings came from a laboratory colony and observation arena, filmed with a fixed 48-megapixel camera at 1920 by 1080 pixels and 30 frames per second under constant LED illumination. The dataset covered 29 sessions: 11 containing trophallaxis and 18 showing routine colony activity. Every fifth frame was exported from sequences with trophallaxis, while every tenth frame was selected from routine-activity sequences. A macro lens was used for close-contact interactions. These choices provided interpretable examples of rare exchanges but also restricted the dataset to one species, one colony setting and standardized imaging conditions; the release does not include continuous raw video in its primary COCO bundle.

Because the material originated as video but the benchmark evaluates still frames, the authors also tested a stricter sequence-safe split in which entire clips remained within one partition. RF-DETR retained an advantage in that complementary evaluation, which used 133 test images and therefore should not be compared directly with the 45-image headline test set. The finding supports the interpretation that the result was not solely an artifact of the track-majority arrangement. Even so, both protocols still score independent frames and do not fully measure the discovery, duration or identity continuity of trophallaxis events. The small test partitions also make the size of the reported metric differences uncertain.

Experiments on an auxiliary dataset containing only densely grouped ants showed a similar pattern. RF-DETR improved detection quality by 0.018 mean average precision over YOLO26n on the dense-ant validation split, but delivered about half the throughput. A separate temporal-processing experiment added ByteTrack association, short-track removal and gap filling to optimized RF-DETR predictions. Smoothing improved matched precision by 0.0161 with essentially no change in headline mean average precision, suggesting that temporal consistency may help stabilize video output without materially improving the underlying frame-level detector.

The authors are explicit that their results establish baselines for frame-level interaction-state detection under controlled laboratory conditions. They do not demonstrate field-ready performance, nutrient-flow measurement, identity-aware interaction networks or full behavioral understanding. Future event-level evaluation is planned around manually verified interaction episodes, temporal intersection-over-union and tracking diagnostics. Within the reported benchmark, however, the practical conclusion is clear: RF-DETR Small offered the strongest detection quality and greatest sensitivity to minority trophallaxis states, while YOLO26n offered the best throughput. The publicly released benchmark provides a common foundation for testing which detector should be used when analysis fidelity or real-time processing is the priority.

Subject of Research: Frame-level detection of normal and trophallaxis states in dense laboratory footage of Camponotus fellah ants

Article Title: A reproducible benchmark for trophallaxis state detection in Camponotus fellah using YOLO26 and RF-DETR

Article References: Kushnir, D. (2026). A reproducible benchmark for trophallaxis state detection in Camponotus fellah using YOLO26 and RF-DETR. Discover Artificial Intelligence, 6(1), Article 1042. https://doi.org/10.1007/s44163-026-01953-2

Image Credits: AI Generated

DOI: 10.1007/s44163-026-01953-2

Keywords: Camponotus fellah, trophallaxis, object detection, RF-DETR, YOLO26, COCO metrics, ant behavior, computer vision, benchmark dataset, transformer detectors, reproducible, benchmark

The study used 613 original photographs collected from mothers who gave birth at the First Hospital of Jilin University in Changchun, China, between March and July 2025. Eligible participants were women aged 18 or older who were lactating and had a nipple injury during their postpartum hospital stay. Two experienced nurses independently assessed each injury using a standardized five-level scale, and a third nurse resolved disagreements. Women were excluded if another dermatological disorder could obscure the injury assessment or if dressings, tattoos, cosmetics or other materials prevented clear observation. The ethics committee approved the protocol, participants provided written informed consent, and the investigators stated that data collection, storage, processing and analysis followed minimization and de-identification procedures. The report did not provide a patient count separate from the number of photographs.

To improve consistency, nurses captured the images with a Fujifilm X-T4 camera fitted with a 60 mm F2.4 R Macro lens. A gray card was used for color calibration, while image stabilization and automatic mode were used to reduce blur and maintain consistent basic settings. The stated acquisition parameters included a shutter speed of 1/60 second, an ISO range of 3,600–6,000, an aperture between F6 and F8, a 60 mm focal length and automatic white balance. Photographs were excluded if they were blurred, overexposed or obstructed by garments, linens or other material that prevented evaluation of the lesion. The grading scale separated stage 0, with an intact surface and normal skin, from stage I, with redness, bruising or swelling despite an intact surface; stage II, involving superficial abrasions, fissures, bleeding, hematoma or ulceration; stage III, involving deep fissures, blisters, deep ulcers or severe erosion; and stage IV, with complete erosion of part of the nipple. Rectangular bounding boxes were added through the Roboflow platform and their labels were subsequently reviewed by a nurse with at least ten years of clinical experience.

The investigators expanded the 613-photograph collection to 2,452 images through online augmentation. All training images were standardized to 640 by 640 pixels and randomly rotated. Modifications to saturation, value and contrast in the HSV color space introduced variation compatible with differences in illumination and color distribution, and random noise was added to improve performance on less clear images. The researchers also introduced task-specific procedures. Lesion-preserving crops were localized more tightly to the nipple for stage 0–II injuries but extended to the areola and breast for stage III–IV lesions. Local contrast enhancement was applied within lesion regions to make early erythema and small fissures more visible, while class-aware sampling increased the probability of including under-represented advanced-stage injuries. The expanded dataset was divided in a 7:2:1 ratio, yielding 1,716 training images, 490 validation images and 245 independent test images. Because these augmented images were derived from the original collection rather than representing additional independent photographs, the test result does not establish performance on a new patient population.

Six single-stage detectors—YOLOX, PICOdet, YOLOv5, RTDETR, YOLOv8 and YOLOv11—were trained for 200 epochs with a batch size of 16 and 640-by-640-pixel inputs. The models used stochastic gradient descent with an initial learning rate of 0.01, weight decay of 0.0005 and momentum of 0.937. Evaluation followed a common COCO-style protocol: a prediction counted as a true positive only when its class matched the reference label and the predicted and reference bounding boxes met the specified intersection-over-union threshold. Precision represented the proportion of predicted injuries that were correct, recall represented the proportion of actual injuries detected, and mAP@0.5 averaged precision across injury classes at an overlap threshold of 0.5. Training, testing and image preparation were conducted with Python 3.8.18, PyTorch 2.1.0 and CUDA 11.8 on a computer with an NVIDIA RTX 4080S graphics card containing 16 GB of memory and an Intel Core i9-13900KF processor.

YOLOv11 led the six-model single-stage comparison. Its precision of 0.85, recall of 0.82, F1 score of 0.83 and mAP of 0.84 were higher than the corresponding values for every alternative. PICOdet ranked second, with precision of 0.78, recall of 0.75, F1 score of 0.76 and mAP of 0.77. YOLOv5 recorded values of 0.75, 0.72, 0.73 and 0.74; YOLOX produced 0.74, 0.71, 0.72 and 0.73; YOLOv8 produced 0.72, 0.69, 0.70 and 0.71; and RTDETR had the lowest single-stage results, with precision of 0.68, recall of 0.65, F1 score of 0.66 and mAP of 0.67. Under the unified mAP@0.5 protocol, the additional baselines were lower: Faster R-CNN reached 0.621, ATSS 0.578, FCOS 0.508, Cascade R-CNN 0.470, CenterNet 0.472 and Libra R-CNN 0.438. For Faster R-CNN and Cascade R-CNN, the corresponding mAP@0.5:0.95 scores were 0.301 and 0.203. These comparisons support YOLOv11 as the best performer within the tested dataset and experimental setup; they do not establish it as the universally best detector for nipple injury or for other clinical image tasks.

The training curves indicated that YOLOv11’s performance rose rapidly and reached a relatively stable plateau after approximately 100 epochs. Its reported recall remained above 0.85 during much of training and sometimes approached 0.9, while precision stayed above 0.85 across most phases and reached 0.9 in some intervals. Distribution-focal, bounding-box and classification losses generally declined and stabilized as training progressed. The confusion analysis reported a recognition rate of 0.96 for normal nipples at stage 0 and 0.87 for stage III injuries. Some stage I and stage IV images were confused, a finding the researchers associated with the visual resemblance of reddened areas in static images. They identified more stage-specific contrast examples, additional texture information and clinical data beyond photographs as possible routes for improving that discrimination. The model’s F1-confidence analysis also showed that overly stringent confidence criteria could sharply reduce the number of accepted detections, illustrating why threshold selection is a performance trade-off rather than a measure of clinical effectiveness.

Ablation experiments indicated that the system’s results depended on several elements rather than architecture alone. YOLOv11’s mAP@0.5 was 0.71 when trained on the 613 unaugmented images, 0.79 with standard geometric and color augmentation, and 0.84 with the stage-aware clinical augmentation pipeline. Resolution also mattered: the score fell to 0.76 at 416 by 416 pixels, reached 0.84 at 640 by 640 and rose only marginally to 0.85 at 800 by 800. Removing distribution focal loss reduced the score to 0.80, disrupting the balance between box and classification losses lowered it to 0.78, and disabling multi-scale detection produced 0.80. The ablation analysis associated the larger decline in stage I–II lesions with the loss of multi-scale feature use. Taken together, these experiments identified augmentation, input resolution, compound loss configuration and multi-scale feature use as contributors to performance. The 640-by-640 setting was described as the best accuracy-efficiency trade-off, while the 800-by-800 setting imposed a higher computational cost for only a small additional gain.

The team used the Ultralytics Hub platform to prepare the optimized YOLOv11 model for distribution through a mobile application, presenting the tool as a possible real-time aid for nurses and patients. Example inference images showed bounding boxes labeled from stage 0 through stage IV, with confidence values from 0.8 to 1.0. The authors argued that such a system could make grading more objective, quantitative and reproducible while potentially reducing the time required for visual assessment. However, the reported work constitutes technical validation rather than a prospective clinical trial of the mobile program. It did not measure whether application-assisted assessment changes nurses’ decisions, reduces assessment time, improves agreement with experts or affects breastfeeding duration. The tool also relied on visual information alone and did not incorporate inputs such as breastfeeding frequency, pain scores, physical examination findings or other multimodal clinical data.

Several limitations constrain how far the findings can be generalized. All photographs came from one hospital and were acquired under standardized conditions by trained investigators, so performance may differ for patients using ordinary cameras, variable lighting, difficult angles or different skin tones. The original sample consisted of 613 photographs without a reported patient denominator, and augmentation increased the image count without adding independent participants. The grading labels were also derived from expert visual assessment, making their consistency relevant to the reference standard. Most importantly, the study evaluated image-based detection and staging against those labels rather than patient benefit or clinical workflow performance. The authors identified multicenter validation, integration of additional clinical information and prospective efficacy evaluations as priorities for future work. Until those steps are completed, the strongest conclusion supported by the research is that YOLOv11 outperformed the tested alternatives on a controlled, single-center nipple-injury image dataset; it remains a potential decision-support candidate rather than a proven replacement for clinical assessment.

Subject of Research: Comparative benchmarking of object-detection models for image-based nipple-injury detection and staging

Article Title: Comparative examination of YOLO models for classification and utilization in nipple injuries

Article References: Zhang, R., Shu, Y., Ding, S., & Gao, Y. (2026). Comparative examination of YOLO models for classification and utilization in nipple injuries. Frontiers in Medicine, 13, Article 1776068. https://doi.org/10.3389/fmed.2026.1776068

Image Credits: AI Generated

DOI: 10.3389/fmed.2026.1776068

Keywords: YOLOv11, nipple injury, medical image analysis, object detection, deep learning, data augmentation, breastfeeding, clinical decision support, single-center study, Comparative, examination, YOLO

The researchers constructed ChildArt-Open from the Google Open Images dataset. Three researchers in children’s art education independently verified the selected images, achieving a Cohen’s κ of 0.86. From 12,400 images identified through the initial screening, the researchers removed duplicates, low-resolution images, non-drawing photographs, adult artworks, digitally synthesized or AI-generated images, and images containing identifiable personal information. The final set comprised 5,000 drawings divided at the image level into 3,500 training, 750 validation and 750 test images. Its categories included people, animals, family scenes, nature, vehicles, fantasy characters and abstract doodles.

DETR served as the spatial backbone, treating object detection as a set-prediction problem within an encoder-decoder Transformer. CLIP supplied image and text representations from a shared semantic space. The researchers inserted a cross-modal attention module after self-attention in each of DETR’s six decoder layers. Twenty CLIP prompt templates were embedded to update object-query features alongside semantic representations. The CLIP ViT-B/32 encoders were frozen, the DETR branch was fine-tuned from COCO weights, and the new projection and cross-modal layers were trained from random initialization.

Training combined DETR classification, bounding-box regression and generalized intersection-over-union losses with a bidirectional CLIP contrastive loss, a Smooth L1 loss for creativity regression and weight regularization. Three experts in children’s art education and developmental psychology independently annotated semantic labels and rated creativity from 1 to 5 across originality, imagination, expressiveness, composition and emotion. Reported agreement measures included Fleiss’ κ of 0.82, Krippendorff’s α of 0.84 and an intraclass correlation of 0.88. The model was trained for 200 epochs over 72 hours using 32 NVIDIA A100 GPUs, AdamW, an initial learning rate of 10^-4 and FP16 mixed precision.

On the 750-image ChildArt-Open test set, DETR-CLIP reached 78.6% ± 0.4% mAP@0.5, compared with 62.3% ± 0.6% for DETR, 71.5% ± 0.5% for Deformable DETR, 73.8% ± 0.6% for Faster R-CNN with CLIP, 74.2% ± 0.5% for RT-DETR, 75.1% ± 0.4% for DINO-DETR and 72.6% ± 0.6% for YOLOv8. The comparison with DETR had p < 0.01 and a Cohen’s d of 1.32. Semantic-understanding accuracy was 87.3% ± 0.5%, with a macro-F1 of 0.86; Faster R-CNN with CLIP had an accuracy of 76.5% ± 0.7%.

For continuous creativity scores, the model had a Spearman correlation of 0.81 ± 0.02 with expert ratings, an intraclass correlation of 0.79 ± 0.03, a mean absolute error of 0.34 ± 0.02 and an RMSE of 0.46. Thresholds determined on the validation set divided creativity into low, medium and high categories. Test-set classification accuracy was 85.5%, with a macro-F1 of 0.84. Ablation experiments reported that removing the cross-modal interaction module was associated with a 15.1% decrease in creativity-scoring accuracy, while removing CLIP semantic guidance was associated with a 9.1% decrease in semantic-understanding accuracy. These ablation results concern the stated evaluation components and do not by themselves establish improved metaphor or creativity understanding from detection gains alone.

Zero-shot semantic interpretation was tested on five concepts not seen during training: loneliness, dream, fear, family warmth and freedom. Using the prompt “a child’s drawing expressing [concept],” DETR-CLIP achieved 78.4% top-1 accuracy, 89.6% top-3 accuracy, 72.8% mAP and a semantic-consistency score of 0.83. Its concept-specific top-1 results ranged from 74.8% for fear to 82.1% for family warmth. CLIP zero-shot classification reached 61.2% top-1 accuracy, while an ordinary object-recognition model reached 52.1% top-1 accuracy.

Among 95 test images that the authors classified as incorrect semantic interpretations, 40 cases, or 42.1%, were attributed to symbolic ambiguity; 29, or 30.5%, to cultural differences; 17, or 17.9%, to multi-object occlusion; and nine, or 9.5%, to other causes. In an external evaluation on the KidsDraw dataset, with model weights fixed and no fine-tuning, the authors reported decreases of 6.7% in mAP@0.5, 8.1% in semantic-understanding accuracy, 9.4% in metaphor-interpretation macro-F1 and 7.5% in creativity-level consistency relative to ChildArt-Open. The results therefore indicated incomplete generalization across differences in drawing style, theme and object-class distribution.

The complete model contained 196.9 million parameters and required 156.2 G floating-point operations. On one NVIDIA A100 GPU with batch size 1 and FP16 precision, measured inference latency was 22.3 milliseconds, equivalent to 44.8 frames per second. Important limitations included the inability to group samples by individual child because the source data contained no identity information, creating a possible data-leakage concern. The study also reported unverified robustness to long text prompts and no explicit modeling of painting media or paper texture. ChildArt-Open identifiers, filtering and annotation files, split information, annotation guidance and model code were released, while the original images remained subject to source copyright restrictions.

Subject of Research: Machine-learning analysis of object detection, semantic interpretation and creativity scoring in children’s drawings

Article Title: A model for children’s drawing content analysis and creativity assessment Integrating DETR target detection and CLIP semantic understanding

Article References: Yuan, Y. (2026). A model for children’s drawing content analysis and creativity assessment Integrating DETR target detection and CLIP semantic understanding. Discover Artificial Intelligence, 6(1), Article 1498. https://doi.org/10.1007/s44163-026-02466-8

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02466-8

Keywords: children’s drawings, DETR, CLIP, object detection, semantic understanding, creativity assessment, zero-shot interpretation, multimodal learning, computer vision, ChildArt-Open, model, children

The researchers developed Fine-Grained Bidirectional Multi-Scale YOLO, or FBMS-YOLO, as a task-driven reconfiguration of YOLOv11 rather than a wholly new detector. Its three main components address different parts of the localization problem. The Attention-enhanced Refinement and Context Network, or ARCNet, models slender and fragmented strokes; the Fine-grained Bidirectional Feature Pyramid Network, or FB-FPN, preserves high-resolution details during cross-scale fusion; and the Dynamic Focusing and Boundary-Penalty Intersection over Union loss, or DFB-IoU, guides bounding-box regression when character edges are ambiguous.

ARCNet replaces the baseline C3k2 feature-extraction blocks with Cross Stage Partial Hierarchical Progressive Multi-scale Feature Aggregation modules, abbreviated CSP-HPMFA. Each module retains a lightweight branch for complementary features and processes another through partial channels with progressive 3×3, 5×5 and 7×7 receptive fields. The authors designed this arrangement to represent short fragments and long slender structures while avoiding the redundant full-channel computation associated with applying large kernels to every feature. The existing AIFI attention component is also used at the deepest semantic stage to model long-range context, including the vertical arrangement of characters and extended material patterns.

FB-FPN replaces direct upsampling-and-concatenation fusion with a Bidirectional Adaptive Fusion Module, or BAFM. Sigmoid gates estimate the reliability of high- and low-level features, allowing semantic information to guide shallow boundary details while low-level information feeds back toward the deeper representation. The framework also extends detection from the original P3–P5 levels to P2–P5, adding a higher-resolution prediction level for weak traces, small fragments and fine character contours. This higher-resolution branch is integrated into bidirectional fusion rather than appended as an independent prediction layer.

For localization supervision, DFB-IoU combines a normalized penalty for differences between the four predicted and ground-truth box boundaries with the dynamic focusing mechanism used in WIoUv2. The dynamic term adjusts each sample’s regression weight according to its localization difficulty relative to an exponential moving average of IoU loss. Unlike CIoU, the proposed loss does not explicitly impose an aspect-ratio consistency term. The authors state that this boundary-based approach is intended for characters whose faded, incomplete or visually entangled edges can make geometric aspect-ratio penalties unstable.

The main DeepJiandu dataset contains 7,416 infrared bamboo slip images with 99,852 annotated character instances across 2,242 categories, divided into 5,922 training, 751 validation and 743 testing images. For this localization study, all annotations were merged into one foreground class. Models were trained for 100 epochs at 640×640 input resolution on a single NVIDIA RTX 5090. The main comparisons and ablations used a fixed seed, while five independent runs with seeds 0 through 4 were used for statistical testing. Welch’s two-sample t-tests found mean mAP50 of 94.32% for FBMS-YOLO versus 92.88% for YOLOv11-n, with p=5.93×10⁻⁶, and mean mAP50:95 of 56.74% versus 54.90%, with p=3.58×10⁻⁷.

FBMS-YOLO used 4.3 million parameters and 17.7 GFLOPs, compared with 41.3 million parameters and 197 GFLOPs for Faster R-CNN. It achieved 94.4% mAP50, 1.3 percentage points above Faster R-CNN’s reported score, while its 56.8% mAP50:95 was 0.1 point below Faster R-CNN’s 56.9%. The accuracy gains came with additional model and computational cost relative to YOLOv11-n, so the authors characterized FBMS-YOLO as an accuracy-oriented detector rather than a lightweight replacement. Its reported inference speed was 71 frames per second, down from 109 frames per second for the baseline in the ablation setup.

Progressive ablation results associated the largest early gains with adaptive fusion and high-resolution prediction. Adding BAFM increased mAP50 from 92.8% to 93.3% and mAP50:95 from 55.0% to 55.8%; adding the P2 level produced 93.7% and 56.0%, respectively. CSP-HPMFA raised those measures to 94.1% and 56.3%, while AIFI further increased mAP50:95 to 56.6%. DFB-IoU then brought the complete model to 94.4% and 56.8% without changing parameter or GFLOP counts. In a separate cross-dataset experiment on 9,306 annotated oracle bone inscription rubbing images, the architecture increased F1 from 83% to 84%, mAP50 from 87.7% to 89.8% and mAP50:95 from 52.5% to 54.1%. The dataset had 8,895 training and 411 validation images but no separate test set, so those results were measured on the validation set.

The study identifies several bounded limitations. Visually connected adjacent characters may be interpreted as one region, and stroke-like cracks or fractured slip edges can still produce false positives. When an annotated boundary is itself ambiguous, the predicted box may cover only the most visible fragments. The oracle bone experiment demonstrated transfer to a different historical character material, but it was limited to bounding-box localization on one dataset and did not test broader domains or fine-grained character recognition. The authors propose future work on instance-level contour modeling, degradation-aware negative learning, structure-preserving losses, anisotropic or deformable convolution, cross-domain learning and joint detection and recognition.

Subject of Research: Benchmark evaluation of an object detector for locating degraded bamboo slip characters

Article Title: A fine-grained bidirectional multi-scale network for extreme-aspect-ratio bamboo slip character detection

Article References: Gao, F., Bao, C., Kong, M., & Chen, B. (2026). A fine-grained bidirectional multi-scale network for extreme-aspect-ratio bamboo slip character detection. Discover Artificial Intelligence, 6(1), Article 1275. https://doi.org/10.1007/s44163-026-01976-9

Image Credits: AI Generated

DOI: 10.1007/s44163-026-01976-9

Keywords: bamboo slip characters, object detection, FBMS-YOLO, YOLOv11, cultural heritage digitization, computer vision, feature fusion, boundary localization, DeepJiandu, fine-grained, bidirectional, multi-scale

The research, published in Discover Artificial Intelligence, used SCB-Dataset5, which contains 7,428 classroom images and 106,830 annotated instances across 20 categories. These included student behaviors such as hand-raising, reading, writing, talking, yawning, phone use and leaning on a desk, together with teacher actions and classroom objects. The images were divided into 5,199 training images, 1,114 validation images and 1,115 test images, and all inputs were resized to 320 by 320 pixels.

The detector combined depthwise separable convolutional blocks, lightweight multi-scale feature fusion and a shallow detection head. Its output tensor contained three candidate boxes at each of 1,600 spatial locations, with each candidate representing four bounding-box coordinates, one objectness score and 20 class scores. The primary federated experiment simulated five clients, each corresponding to a classroom, school or edge node. In the non-independent and identically distributed evaluation, class allocations across those clients were generated with Dirichlet concentration parameters of 1.0, 0.5 and 0.1 to represent mild, moderate and severe heterogeneity.

Each simulated client trained locally and sent updated model parameters to a central server rather than transmitting raw images, labels or bounding-box annotations. Proximal regularization penalized excessive divergence of local updates from the global model. On the main SCB-Dataset5 test set, centralized training produced 98.12% F1-score, 98.42% mAP@0.5 and 86.35% mAP@0.5:0.95. The proximal federated model achieved the same 98.12% F1-score, along with 98.39% mAP@0.5 and 86.21% mAP@0.5:0.95.

Under the non-IID partition, standard FedAvg produced 96.24% F1-score and 96.58% mAP@0.5, compared with 98.12% F1-score and 98.39% mAP@0.5 for the proximal strategy. Performance declined gradually as heterogeneity increased: the proximal approach retained 97.84% F1-score under mild skew, 97.21% under moderate skew and 96.34% under severe skew. In an extended client-scalability simulation, mAP@0.5 decreased from 97.42% with five clients to 96.24% with 30 clients, while estimated communication cost rose from 2.37 MB to 14.22 MB per communication round.

The aggregated global model was then compressed through batch-normalization folding, 40% structured channel pruning, 20 epochs of post-pruning fine-tuning and full-integer INT8 quantization. Weights used per-channel symmetric INT8 quantization, activations used per-tensor asymmetric INT8 quantization, and biases were accumulated in INT32. The final compressed federated model achieved 95.52% F1-score, 95.64% mAP@0.5 and 78.86% mAP@0.5:0.95. Relative to the uncompressed federated model, its storage fell by 69.2% and its measured latency declined by 65.3%.

Embedded tests ran on an NXP FRDM-K64F development board with a 32-bit Arm Cortex-M4F processor, 1 MB of Flash memory and 256 KB of SRAM. The compressed model occupied 228.9 KB, while the complete TensorFlow Lite Micro runtime used 118.3 KB of SRAM for the tensor arena, activation and input/output buffers, and post-processing space. Each 320-by-320 image took 64.8 milliseconds through the full on-device path, including INT8 execution, output decoding, confidence filtering at 0.25 and non-maximum suppression at an IoU threshold of 0.50. Camera acquisition and image transmission were not included in that timing.

Diagnostic privacy tests indicated lower, but not eliminated, exposure relative to centralized full-precision training. Membership-inference attack accuracy fell from 67.84% in centralized FP32 training to 55.92% in the federated TinyML configuration; reconstruction PSNR declined from 21.62 dB to 14.86 dB, and model-inversion success fell from 38.71% to 12.58%. In a stress test with 30% malicious clients, mAP@0.5 was 94.12%. The method did not include secure aggregation, differential privacy, trust weighting or robust aggregation, and no direct board-level power measurements were reported.

Cross-dataset evaluation followed dataset-specific fine-tuning rather than zero-shot testing. The compressed federated INT8 model achieved 93.58% mAP@0.5 on UK_Dataset and 92.28% on SiTBehavior, below its SCB-Dataset5 result. The study was a computational federated simulation rather than a live multi-school deployment, and most evidence came from SCB-Dataset5, with limited validation on two additional datasets. Real-classroom testing, direct energy measurements, heterogeneous-hardware evaluation and stronger privacy and malicious-client defenses remain necessary before wider use.

Subject of Research: Evaluation of a simulated federated TinyML framework for detecting student behaviors in smart classrooms

Article Title: A real-time federated TinyML framework for student behavior detection in smart classrooms

Article References: Yahyati, C., Lamaakal, I., El Makkaoui, K., & Ouahbi, I. (2026). A real-time federated TinyML framework for student behavior detection in smart classrooms. Discover Artificial Intelligence, 6(1), Article 1282. https://doi.org/10.1007/s44163-026-02362-1

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02362-1

Keywords: student behavior detection, federated learning, TinyML, microcontroller deployment, INT8 quantization, model pruning, classroom analytics, privacy-preserving AI, object detection, edge inference, real-time, federated

The work focused on six classes of hot-rolled steel surface defects represented in NEU-DET: crazing, inclusions, patches, pitted surfaces, rolled-in scale and scratches. The dataset contains 1,800 grayscale images at 200×200 pixels, with 300 images per class. Researchers randomly divided the data into 1,440 training, 180 validation and 180 test images, an 8:1:1 split. Images were scaled to 640×640 pixels during training, and all reproduced models used the same data split, training duration and augmentation settings.

EDGI-YOLO was developed from YOLOv8n, which had approximately 3.01 million parameters and 8.2 GFLOPs. Its architecture combines four modules. ECA-GC first screens feature channels with efficient channel attention and then applies global context modulation. DSEConv separates spatial filtering from channel transformation with depthwise separable convolution and adds Squeeze-and-Excitation attention. AllGhostC2f replaces standard C2f blocks with a Ghost-convolution-based structure designed to reduce redundant feature generation. Finally, IAFHead uses DSEConv in its classification and regression branches, fuses information across three feature scales with CrossScaleSE attention and applies different branch depths to shallow and deep features.

The ablation experiments examined every single-module, two-module and three-module combination as well as the complete model. The full four-module configuration reached 78.85% mAP@0.5, compared with 74.91% for the baseline. Combining DSEConv and AllGhostC2f without the other components produced 74.73%, while ECA-GC paired with IAFHead reached 78.03%. The full model’s gain over the baseline was not reproduced by either component alone, indicating that the measured performance depended on the combined design.

Removing one component from the complete system also reduced performance. The configuration without IAFHead achieved 76.97% mAP@0.5, while models without AllGhostC2f, DSEConv or ECA-GC achieved 76.02%, 76.69% and 76.03%, respectively. These reductions ranged from 1.88 to 2.83 percentage points relative to the full model. Separate tests of ECA-GC found that its cascaded form at backbone layer 5 produced 76.42% mAP@0.5, exceeding the standalone ECA, standalone GC, reversed-cascade and parallel arrangements by 0.80, 0.36, 0.37 and 1.34 percentage points, respectively.

In comparisons with other detectors, EDGI-YOLO recorded the highest overall mAP@0.5 and mAP from 0.5 to 0.95 among the models evaluated, with scores of 78.85% and 44.70%. Its precision was 79.52%, while recall was 71.54%. Class-level AP@0.5 included 58.05% for crazing, 90.36% for patches, 67.81% for rolled-in scale and 97.06% for scratches. Compared with YOLOv8n, crazing increased from 46.5% to 58.05%, and rolled-in scale from 51.9% to 67.81%.

Results from five training seeds gave mean mAP@0.5 values of 77.98% ± 0.92% for EDGI-YOLO and 75.12% ± 1.00% for YOLOv8n. EDGI-YOLO produced a higher score with each of the five seeds, with improvements ranging from 0.30 to 4.46 percentage points. The study therefore reported that its advantage over the baseline was not confined to one initialization, although the article does not provide a separate statistical hypothesis test.

Deployment was evaluated on a system with an 11th-generation Intel Core i5-11400H processor and an NVIDIA GeForce RTX 3050 Laptop GPU with 4 GB of memory. With TensorRT FP16, EDGI-YOLO processed 159.26 frames per second at 6.28 milliseconds per image, compared with 146.68 frames per second for YOLOv8n, and retained a reported 4.27-percentage-point mAP@0.5 advantage. In the ONNX CPU configuration, it ran at 27.16 frames per second with 36.82 milliseconds per image, versus 24.87 frames per second for YOLOv8n. Reported file sizes were smaller in PyTorch and ONNX, but the study did not separately measure loading time or memory use.

Cross-dataset testing used 3,570 aluminum surface images in GC10-DET, covering 10 defect categories and an 8:1:1 split. EDGI-YOLO achieved 64.29% mAP@0.5, 4.38 percentage points above YOLOv8n, while retaining 1.66 million parameters and 6.05 GFLOPs. The authors reported improvements of 2.44 percentage points for crease, 1.16 for inclusion and 3.90 for rolled pit. The evidence remains limited to two public image datasets, selected comparison models and the specified inference hardware; it does not establish performance on every industrial surface, production environment or defect type.

Subject of Research: Lightweight neural-network detection of steel and aluminum surface defects in image-based benchmarks

Article Title: A lightweight model for steel surface defect detection based on EDGI-YOLO

Article References: Wu, H., & Cao, F. (2026). A lightweight model for steel surface defect detection based on EDGI-YOLO. Discover Artificial Intelligence, 6(1), Article 1280. https://doi.org/10.1007/s44163-026-02329-2

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02329-2

Keywords: EDGI-YOLO, steel surface defects, YOLOv8n, computer vision, object detection, NEU-DET, GC10-DET, edge deployment, TensorRT, lightweight, model, steel

This was a computational model-development and benchmarking study rather than a review or clinical investigation. Its scope included pig-face images from the PetFace pig subset and the FIRC dataset, which the authors divided into training, validation and test sets at a ratio of 7:2:1. Data augmentation included random flipping, random cropping and color-gamut transformation to represent variations such as backlighting, low light, occlusion and changes in posture. All comparison models used fivefold cross-validation, and the researchers used an independent-samples t test to assess differences between groups.

On the PetFace data, I-YOLOv8-MCGIRNet recorded precision of 0.956, recall of 0.948 and mAP@0.5:0.95 of 0.897. The reported p-values for comparisons with the other models were below 0.05. In separate performance-and-efficiency comparisons, the model used 15 million parameters and produced mAP@0.5 of 0.966 on PetFace. Comparative results reported for that dataset included approximately 0.933 for YOLOv8s-MCGIRNet, 0.947 for I-YOLOv8-Swin Transformer and 0.885 for Faster R-CNN, whose model size was about 68 million parameters.

The FIRC results showed a similar pattern. I-YOLOv8-MCGIRNet achieved mAP@0.5 of 0.955 with 15 million parameters. The reported mAP@0.5 values were 0.924 for YOLOv8s-MCGIRNet, 0.937 for I-YOLOv8-Swin Transformer and 0.875 for Faster R-CNN. An ablation experiment attributed successive gains to the global attention module, BiFPN feature fusion, CIoU loss and MCGIRNet. Embedding MCGIRNet increased mAP@0.5 to 0.966 while reducing the model to 15 million parameters, according to the study.

Under the researchers’ definition of recognition consistency, the model’s value increased from about 90.9% to 94.5% in normal lighting, from 87.9% to 94.8% in dark light below 200 lux and from 85.4% to 92.7% under backlighting with a light-intensity difference above 800 lux. The computational tests reported 16.32 GFLOPs for I-YOLOv8-MCGIRNet, compared with 19.68 for YOLOv8s-MCGIRNet, 22.75 for I-YOLOv8-IRNet, 47.21 for I-YOLOv8-Swin Transformer and 118.46 for Faster R-CNN. Peak GPU memory was 8.26 GB on PetFace and 7.94 GB on FIRC, while reported epoch times were 4.12 and 3.28 minutes, respectively.

Evaluation in large-scale pig-farm settings covered nursery, fattening and breeding houses. Each scene type included 10 breeding stalls, with pig images collected over 30 consecutive days under varying lighting, postures and stocking densities. I-YOLOv8-MCGIRNet was reported to detect one pig face in 2.8 milliseconds. Its small-target detection rate in dense scenes was 94.1%, compared with 86.3% for I-YOLOv8-IRNet. When 30% of local facial features were partially missing, its recognition accuracy was 82.7%, versus 79.6% for I-YOLOv8-Swin Transformer and 61.8% for Faster R-CNN.

Counting performance was measured in sparse, dense and overlapping scenarios. For I-YOLOv8-MCGIRNet, counting accuracy rose from 88.6% to 98.5% in sparse scenes, from 79.9% to 94.7% in dense scenes and from approximately 71.8% to 92.2% in overlapping scenes. In the overlapping condition, Faster R-CNN rose from about 52.1% to 64.9%. These results apply to the sampled farm environments and test conditions used in the study rather than establishing performance across every type of pig facility.

Recognition was also assessed with pigs standing, lying down and turning their heads 45 degrees left or right. The study reported that I-YOLOv8-MCGIRNet achieved the best detection and recognition accuracies among the compared models for all three postures, with both measures remaining above 0.91 in the non-standard postures. In the head-turning condition, its detection and recognition accuracies were 0.938 and 0.919. Relative to I-YOLOv8-IRNet, the reported improvements were 7.0 and 9.4 percentage points, respectively.

The evidence supports the model’s measured performance within the selected datasets and farm scenarios, but its generalizability remains untested. The authors trained, validated and tested within PetFace and FIRC separately and did not perform direct cross-dataset transfer validation. The study also did not fully evaluate outdoor conditions in extreme weather or optimize the system specifically for facial differences among pig breeds and ages. The results therefore indicate benchmark and scenario-specific performance for pig-face recognition and counting, rather than demonstrated effectiveness in unstudied farms, species, breeds or environmental conditions.

Subject of Research: Experimental development and benchmarking of a pig-face recognition and counting model

Article Title: Intelligent recognition and counting of pig faces based on improved YOLOv8 and MCGIRNet

Article References: Kang, J., Wang, H., Chen, C., & Liu, Q. (2026). Intelligent recognition and counting of pig faces based on improved YOLOv8 and MCGIRNet. Discover Artificial Intelligence, 6(1), Article 1467. https://doi.org/10.1007/s44163-026-02349-y

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02349-y

Keywords: pig-face recognition, pig counting, YOLOv8, MCGIRNet, computer vision, object detection, precision livestock, smart farming, machine learning, Intelligent, recognition, counting