The new framework, called CDCNet, was developed by Yujie Shen and colleagues at the People’s Public Security University of China, the Institute of Forensic Science of China, and Shanxi University. Writing in Frontiers in Bioengineering and Biotechnology, the team reports that their model reaches a mean Average Precision of 83.2 percent on a purpose-built dataset of high-resolution ink footprints, an improvement of 5.2 percentage points over the YOLOv10m baseline it is built upon. The work is framed not as a replacement for forensic examiners but as a step toward intelligent, automated forensic identification.

The difficulty of footprint analysis lies in its scale. A single footprint image contains structures that vary enormously in size and morphological subtlety: a big toe may be long or short relative to its neighbors, an arch may be high, narrow, or flat, and a heel may be round, oval, or rectangular. These distinctions carry biometric significance, yet they are subtle enough that generic object detectors often blur or misclassify them. Traditional approaches relied on manually designed features such as contour extraction and geometric ratios, which could capture local shape but missed the structural relationships linking toes, arch, and heel within the whole footprint.

Deep learning promised to change that, but early convolutional networks faced their own limits. Standard convolutions aggregate information within a regular receptive field, which is poorly suited to the elongated and curved boundaries that define footprint regions. Multi-scale features—those captured at different resolutions of the image—also tend to be fused clumsily, so fine details like toe edges can be washed out while the network processes the broader geometry of the foot. The Chinese team’s answer was to redesign the detection pipeline in three complementary directions: context, orientation, and semantics.

The first pillar is a context-guided spatial feature reconstruction pyramid network, or CGSFRPN. At its core sits a pyramid context extraction module that gathers feature maps from three processing depths of the network—corresponding to one-eighth, one-sixteenth, and one-thirty-second of the input resolution—and pools them down to a shared scale, producing a unified multi-scale summary of the footprint. This pooled pyramid then passes through stacked rectangular self-calibration modules, which use horizontal and vertical pooling to build rectangular attention regions, calibrate them with large strip-shaped convolution kernels, and re-weight the original features. The effect is a network that understands not just what it sees locally but how each patch relates to the footprint’s overall axis and layout.

The second pillar tackles direction. Toe boundaries, arch contours, and heel outlines are elongated or curved along specific orientations, and a conventional square convolution kernel responds weakly to such directional structure. The team’s Direction-Aware Pinwheel Convolution, DAPConv, deploys four branches of asymmetric 1-by-3 and 3-by-1 kernels with carefully chosen asymmetric padding, so that each branch specializes in one spatial direction. The four directional outputs are concatenated and compressed with a small convolution, feeding orientation-sensitive boundary information directly into the backbone of the detector. This allows CDCNet to distinguish regions that look globally similar but differ in the curvature or tilt of their edges—exactly the kind of variation that separates a flat arch from a narrow one.

The third pillar, called ConvFormer, replaces several standard feature extraction blocks in the YOLOv10 architecture with a design rooted in the MetaFormer concept. ConvFormer combines depthwise separable convolutions, which process spatial and channel information efficiently, with a channel-wise multilayer perceptron and a custom StarReLU activation. The result is efficient local-to-global semantic modeling: fine structures in the forefoot retain their detail, while broader contextual relationships across the midfoot and rearfoot are reinforced. Ablation experiments showed each pillar contributes, and that the fully assembled model improves both strict and lenient detection metrics while actually lowering computational cost from 58.9 to 54.6 gigafloating-point operations per second relative to the baseline.

The experiments rested on a carefully constructed dataset of 2,587 high-resolution ink-print footprints collected from 573 individuals. Ink was applied to the plantar surface with a sponge applicator and impressions were made on absorbent Xuan paper on a hard, flat indoor surface; blurred or incomplete samples were excluded or recollected, and the paper impressions were digitized with a high-resolution scanner. Experienced annotators labeled eight fine-grained categories: thumb-toe length and second-toe length in the forefoot, high, narrow, and flat arch in the midfoot, and round, oval, and rectangular heel in the rearfoot. Crucially, the data were split into training, validation, and test sets at the subject level, so no individual’s footprints appeared in more than one subset—a safeguard against identity leakage that makes the reported generalization more credible.

Head-to-head comparisons under identical training conditions put CDCNet ahead of Faster R-CNN, YOLOv8, YOLOv10, YOLOv12, RT-DETR, and Mamba-YOLO. The classic two-stage Faster R-CNN managed only 48.13 percent mean Average Precision, while the strongest rivals, YOLOv12 and Mamba-YOLO, reached 80.4 and 81.2 percent respectively. Visual inspections highlighted the difference: competing detectors missed long thumbs and oval heels, mislabeled narrow arches as flat, and produced overlapping, unstable boxes when the confidence threshold was lowered to 0.3. CDCNet, by contrast, kept its detections clean and correctly placed at both thresholds, with no missed regions or category confusion in the reported examples.

The authors acknowledge that their results come from a proprietary dataset, and future work will aim to strengthen generalization across acquisition scenarios, deploy the model on edge devices in complex field conditions, and broaden applications in related domains. Even so, the study marks a notable convergence of computer vision and forensic trace examination: a system that treats a footprint not as a single image to classify, but as an anatomical composition of interacting regions. If such tools mature, the humble footprint—dusty, partial, or inked on paper—could become an even sharper instrument for narrowing investigations and identifying suspects.

Subject of Research: Deep learning-based fine-grained barefoot footprint region recognition for forensic identification

Article Title: Contextual and structural representation learning for fine-grained barefoot footprint region recognition

Article References: Shen, Y., Xie, W., Zhao, Y., Jiang, X., & Wang, J. (2026). Contextual and structural representation learning for fine-grained barefoot footprint region recognition. Frontiers in Bioengineering and Biotechnology, 14, Article 1912110. https://doi.org/10.3389/fbioe.2026.1912110

Image Credits: AI Generated

DOI: 10.3389/fbioe.2026.1912110

Keywords: barefoot footprint recognition, forensic science, CDCNet, YOLOv10, object detection, computer vision, CGSFRPN, DAPConv, ConvFormer, biometrics, multi-scale feature fusion, deep learning

The system, called Spitz, was developed by a team of forensic scientists at the Civil Police of the Brazilian Federal District in Brasília, working with real casework material rather than curated laboratory samples. At its core is a YOLOv10-L deep neural network, a member of the You Only Look Once family of object detectors that has become one of the most widely used architectures in real-time computer vision. Unlike slower two-stage detectors that first propose candidate regions and then classify them, YOLO models process an entire image in a single pass, predicting bounding boxes and class probabilities simultaneously. That speed matters in forensic settings, where a single microscope slide can contain thousands of fields of view and laboratories face persistent backlogs.

Training the model required a dataset grounded in the messy reality of forensic practice. The researchers assembled 399 microscopic image patches drawn from actual sexual assault cases, all stained with the Christmas Tree stain, the red-and-green histochemical preparation that is a staple of forensic sperm identification. This choice of training data is significant. Staining artifacts, epithelial cells, debris, and variable staining intensity make real casework slides far harder to interpret than idealized reference images, and models trained only on clean laboratory data often falter when confronted with authentic evidence. By anchoring the training set in genuine case material, the Brasília team aimed to build a detector that would hold up under the conditions examiners actually face.

The performance figures reported for the detection model are impressive, though they also reveal the inherent difficulty of the task. Spitz achieved a precision of 94 percent, meaning that when it flags an object as a spermatozoon, it is almost always correct. Recall, the proportion of true sperm cells it manages to find, came in at 72.3 percent, and the mean average precision at a 50 percent intersection-over-union threshold reached 0.842. In practical terms, the system errs on the side of caution: it rarely cries wolf, but it does miss some cells. For forensic work, that trade-off is arguably the right one, because the system is designed not to replace the examiner but to guide them, and a human expert remains in the loop for final verification.

What elevates Spitz beyond a standalone detector is the workflow built around it. The model was integrated into a web interface that supports batch inference across entire slides, reconstructs the spatial layout of the microscope slide, and records the coordinates of every detected sperm cell. An examiner can then navigate the microscope directly to flagged locations for confirmation, rather than sweeping the slide blindly. This coordinate-based verification bridges the gap between automated image analysis and the downstream steps of forensic genetics, where precise localization of sperm cells can feed into techniques such as laser capture microdissection and sperm cell sorting for DNA profiling. The system effectively turns the microscope from a search instrument into a confirmation instrument.

To test whether the technology actually improves forensic practice, the team ran a blind comparative study with three experienced forensic experts who re-examined 30 sexual assault samples using both conventional manual examination and the AI-assisted approach. For each sample, the examiners recorded spermatozoa counts, assigned a classification of Not present, Rare, or Occasional, and logged examination time. The design directly targeted one of the most stubborn problems in forensic biology: inter-examiner variability. Previous research has documented substantial differences in how analysts assess evidence suitability and interpret the same slides, and such variability can affect whether a sample proceeds to DNA extraction at all.

The statistical results favored the AI-assisted workflow across nearly every measure. AI-assisted detection found significantly more spermatozoa than manual examination, and the counts showed lower variability between examiners. Inter-rater agreement, quantified with Fleiss’ Kappa, jumped from 0.698 under manual examination to 0.858 with AI assistance, while raw agreement rose from 73.3 percent to 86.7 percent. The coefficient of variation across classification categories dropped from a mean of 93.5 percent to 80.1 percent, indicating that examiners converged far more closely on how to categorize samples when the algorithm had already flagged candidate cells. Wilcoxon signed-rank tests underpinned these comparisons, lending statistical weight to what the numbers show.

Examination time told a more nuanced story. AI-assisted examinations averaged 6.39 minutes with a standard deviation of 0.59 minutes, compared with 6.93 plus or minus 0.70 minutes for manual examination, a difference that did not reach statistical significance with a p-value of 0.2621. The modest time saving suggests that the primary benefit of Spitz is not speed but quality: more sensitive detection and more reproducible conclusions. The authors also highlight a benefit that is easy to overlook in the statistics: reduced examiner exposure to the intense illumination of the microscope, an occupational burden for analysts who spend their careers hunched over eyepieces.

The study arrives amid a broader wave of machine learning adoption in forensic science. Earlier efforts explored deep convolutional networks for sperm detection on microscope slides, YOLOv5-based sperm cell detection, and automated detection for laser capture microdissection, while commercial platforms have begun offering AI-assisted sperm finding. At the same time, critical reviews of machine learning in forensic DNA profiling have cautioned that validation standards must be rigorous, that datasets are often sensitive and hard to share, and that algorithms must earn the trust of courts and examiners alike. Spitz contributes to this conversation by reporting a validation study built on real casework, blind comparison, and established statistical measures rather than laboratory-only benchmarks.

There are, of course, limits to what the current study demonstrates. The annotated dataset and model weights cannot be released publicly because the underlying images come from sensitive forensic casework protected by institutional data policy, although the code for the Spitz interface may be shared on reasonable request. The recall of 72.3 percent means the algorithm alone is not exhaustive, which is precisely why the system is framed as an examiner’s assistant rather than an autonomous judge. Still, the direction of travel is clear. A model trained on 399 image patches from genuine cases, wrapped in a workflow that respects the expertise of human analysts, outperformed those same analysts on sensitivity and consistency in a blind trial. For a field where the stakes are measured in justice delivered or denied, that combination of technological ambition and procedural humility may prove to be the most important finding of all.

Subject of Research: Deep learning-based automated detection of spermatozoa in forensic microscopy for sexual assault investigations

Article Title: Spitz: a YOLO-based deep learning system for automated spermatozoa detection in forensic genetic analysis

Article References: Humanes, A. C., Calil, A. L. A., Wawruk, H. D., Carneiro, L. D., Silveira, K. B., Yen, W. C., de Andrade Gomes, J., & Meirelles, A. L. S. (2026). Spitz: a YOLO-based deep learning system for automated spermatozoa detection in forensic genetic analysis. International Journal of Legal Medicine. https://doi.org/10.1007/s00414-026-04013-7

Image Credits: AI Generated

DOI: 10.1007/s00414-026-04013-7

Keywords: forensic genetics, deep learning, YOLOv10, spermatozoa detection, sexual assault evidence, automated microscopy, inter-rater agreement, Christmas Tree stain, object detection, DNA profiling, forensic science, computer vision

The core problem Li set out to solve is a familiar one in robotics: the tension between accuracy and speed. Traditional vision models that detect and locate objects well tend to be computationally heavy, making them difficult to deploy on the edge devices that sit next to robot arms on production lines. Earlier approaches, from sliding-window detectors with hand-crafted features to early convolutional neural networks, either generalized poorly or demanded more processing power than an embedded system could spare. Prior YOLO-based grasping systems improved matters, but studies cited in the paper show they still struggled with high resource overhead, degraded accuracy in dark or occluded scenes, and difficulty adapting to complex, changeable environments. Li’s answer is a two-part architecture: an improved, slimmed-down YOLOv10 for detecting targets, paired with a generative grasping convolutional network that refines the robot’s actual grip posture.

The detection side of the system is where most of the engineering ingenuity lies. Li replaced YOLOv10’s original backbone with FasterNet, a network built around Partial Convolution, or PConv, which applies convolution only to a subset of a feature map’s channels rather than all of them. The mathematics is elegant: because floating-point operations scale with the square of the channel ratio, restricting convolution to one quarter of the channels cuts the computation to exactly one sixteenth of a standard convolution. To make that saving task-aware, Li added a channel attention mechanism that reweights channel responses according to the depth gradient distribution of the scene, prioritizing channels carrying object-surface and edge information while suppressing flat background regions. The result is a backbone that spends its limited compute budget where it matters most.

Lightweighting a network usually costs accuracy, so the design compensates in the neck of the network, where features are fused across scales. Li introduced a Context Guided Block, which combines local and atrous convolution to enlarge the effective receptive field and sharpen the capture of occluded targets, and a dynamic upsampling module called DySample, which generates a dynamic point-sampling set to restore spatial detail lost by the compressed backbone. The detection head also gained a rotated bounding box, parameterized as a five-dimensional vector including an angle, together with a dedicated angle-prediction branch trained with Smooth L1 loss. This matters because industrial components tilt and stack; an axis-aligned box cannot describe a screw lying at forty-five degrees, and the resulting positioning errors cascade into failed grasps. A combined loss function integrating bounding box, classification, confidence and angle terms trains the whole system to be simultaneously accurate in position, class, confidence and orientation.

The second half of the framework addresses what happens after detection: deciding exactly how the gripper should approach the object. For this, Li turned to the Generative Grasping Convolutional Network, or GGCNN, and upgraded it with two modules. An improved Inception block extracts multi-scale spatial features in parallel through branching convolutions, using stacked pairs of three-by-three convolutions in place of five-by-five kernels to preserve the receptive field while trimming computation. A Weighted Feature Fusion module then tackles a subtle but important problem: shallow features carry precise edge localization but high-frequency noise, while deep features carry semantic meaning but spatial offset. Directly adding them produces misaligned features, so the WFF module learns trainable channel weights that minimize the alignment error between the two feature spaces, filtering out contradictory noise. The network outputs, for every pixel, a grasp success probability, a gripper rotation angle between minus ninety and plus ninety degrees, and an opening width, from which the system selects the highest-confidence grasp and maps it into the robot’s world coordinates.

The hardware pipeline behind the software is equally concrete. An Intel RealSense D455 depth camera captures synchronized color and depth streams, with color at 1280 by 800 pixels at 30 frames per second and depth at 1280 by 720 at 90 frames per second, covering a detection range from 0.4 to 10 meters. Pinhole imaging mathematics converts three-dimensional world points into two-dimensional image coordinates, and hand-eye calibration establishes the transformation between the camera frame and the base of a UR5 six-degree-of-freedom collaborative manipulator, modeled through Denavit-Hartenberg parameters. A priority scoring model weighs each candidate object’s depth distance and the gap between its center of gravity and the proposed grasp point, so the robot grabs the nearest, most stable targets first. Each grasp is described by four parameters: the projected center position, the gripper rotation angle, and the pre-opening width converted to physical units through the depth data.

The experimental results are striking. Trained and tested on a self-built dataset of 15,000 images spanning 423 industrial and everyday object types, plus the public GraspNet-1Billion benchmark with its billion-scale grasp annotations, the improved YOLOv10 achieved a precision of 0.971, a recall of 0.955 and a mean average precision at the 0.5 intersection-over-union threshold of 0.942. That mAP50 figure comfortably outperformed standard YOLOv10 at 0.918, YOLOv8 at 0.909 and Faster R-CNN at 0.875. Against dedicated grasping systems on GraspNet-1Billion, the method exceeded the GraspNet baseline by 8.0 percent, GG-CNN by 10.1 percent and a YOLOv8-based detector by 3.3 percent, while running at 192 frames per second with only 26.8 gigafloating-point operations, a computation reduction of 82.9 percent compared with GraspNet and 60.8 percent compared with GG-CNN. Ablation tests confirmed the division of labor: FasterNet drove the efficiency gains, while the CG-Block and DySample modules recovered and then surpassed the baseline accuracy, lifting mAP50 by 0.027 over the unmodified network.

Robustness under adverse conditions may be the most eye-catching result. In multi-scenario tests, the model detected 100 percent of target objects in dim lighting, in occluded scenes, in multi-background settings and among visually similar items, where comparison models faltered. In one dark test scene, Faster R-CNN could barely identify two people in a corner while the improved detector found all six people in a bus with the highest accuracy. In occlusion tests with vehicles, it identified five cars where Faster R-CNN managed three. Channel pruning added further efficiency: at a pruning rate of 0.4, the model reached its best mAP50 of 0.937 while shrinking its parameter count and accelerating inference. In a 100-minute real-robot experiment across three scenarios, from metal parts larger than three centimeters to everyday objects to screws and electronic components smaller than three centimeters, posture optimization via the improved GGCNN raised the detector’s mAP50 in the first scenario from 0.850 to 0.921, with similar gains for the comparison models. Across ten complex scenarios combining dimness, occlusion and dust, the full system achieved a minimum grasping success rate of 92.3 percent and held positioning angle deviation within minus 0.04 degrees, far tighter than the 0.23 and 0.27 degree deviations of YOLOv8 and Faster R-CNN.

The improved GGCNN itself earned its place through ablation. On GraspNet-1Billion, the full model with Inception and WFF modules reached an area under the curve of 0.951, versus 0.892 for GGCNN with WFF alone, 0.883 with Inception alone and 0.841 for the baseline network. Detailed validation showed the baseline GGCNN’s grasp angle error of 6.2 degrees falling to 4.5 degrees with Inception, then to 2.3 degrees with WFF added, while the grasp success rate climbed to 95.1 percent with only a marginal increase in inference time. These numbers matter against a backdrop of staggering industrial scale: the paper notes that China alone had more than two million industrial robots in operation in 2024, over half the world’s total, with a manufacturing robot density of 392 units per ten thousand people. Even small gains in grasping reliability translate into enormous productivity differences at that scale.

Li is candid about the limits. The improved YOLOv10 struggles with extremely small targets under one centimeter, and real-time performance can fluctuate under extreme interference during edge deployment. Future work, the paper suggests, could add super-resolution reconstruction modules to enhance micro-target features and further optimize the lightweight architecture for smoother edge inference. Even so, the study demonstrates a coherent design philosophy: rather than piling modules together, it couples a mathematically grounded efficiency mechanism in the backbone with context-aware feature repair in the neck and angle-aware detection at the head, then hands off to a posture network that bridges the semantic gap between edges and meaning. For factories racing toward Industry 4.0, that combination of precision, speed and resilience in the face of darkness, clutter and occlusion could be exactly what robots need to finally grasp the world as reliably as they move through it.

Subject of Research: Lightweight deep learning for robotic grasping detection and grasp posture optimization

Article Title: Robot grabbing detection based on lightweight YOLOv10 algorithm

Article References: Li, H. (2026). Robot grabbing detection based on lightweight YOLOv10 algorithm. Discover Artificial Intelligence, 6(1), Article 1310. https://doi.org/10.1007/s44163-026-02228-6

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02228-6

Keywords: robotics, YOLOv10, object detection, robotic grasping, GGCNN, lightweight neural networks, computer vision, industrial automation, deep learning, grasp posture estimation, edge deployment, intelligent manufacturing