The prototype's layout mirrors the sequence of decisions an apple must undergo. Defect detection and upper-layer ejection come first; size grading is handled below. Because the entire vision pipeline runs on one camera covering three lanes, the design avoids per-lane cameras and keeps the compute budget small enough for real-time deployment.

The authors describe the system in the journal Artificial Intelligence in Agriculture, positioning it as a perception and decision framework rather than a standalone detector: the same trajectory information that stabilizes tracking also drives the ejection logic. At the core of the detector sits MOD-DETR, a network built on the RT-DETR architecture. The team streamlined the convolution designs, added multi-scale context modeling, and kept the model lightweight.

On their apple defect dataset, MOD-DETR reached a mean average precision at an IoU threshold of 0.5 (mAP@50) of 86.7 percent, with precision of 90.7 percent and an F1-score of 85.7, while requiring 28.9 GFLOPs and running at 95.4 frames per second on an RTX 4090 GPU. The authors argue that this balance between accuracy and model complexity matters more than raw speed for deployment on modest hardware, where a heavier network would either miss the real-time budget or sacrifice accuracy. Baselines trained under identical settings put the trade-off in perspective.

YOLOv8s ran faster at 230.8 FPS but scored 5.3 percentage points lower on mAP@50, while YOLOv11m matched accuracy only at more than double the computational cost. Across three random seeds, the defect-only mean AP@50 rose from 80.36 to 82.29 percent, with bruise improving the most at 2.85 percentage points. The largest gains appeared for defects that are easily overshadowed by fruit-surface texture, which the streamlined designs and multi-scale context modeling were intended to address.

Detection alone cannot eject an apple; the system must know which detections belong to which trajectory. For this, the team built ASORT, a tracker based on the SORT framework. Because apples on the line look highly similar and their motion is dictated by the fruit-cup conveying mechanism, complex appearance modeling and re-identification features offer little benefit.

ASORT instead uses simplified state estimation with a constant-velocity motion model, filters apple detections at a confidence threshold of 0.9 while retaining weaker defect detections above 0.3, and associates frames by fusing box overlap with a velocity-consistency score computed from the cosine similarity between current displacement and historical trajectory velocity. Trajectory lifecycle management confirms a track only after a minimum number of consecutive frames and removes it after a maximum number of unreachable frames. On the annotated sorting video, ASORT achieved 97.6 percent MOTA and 98.8 percent IDF1 at 19.7 FPS, cutting false negatives from 101 to 46 and fragmentation from 9 to 4 relative to SORT, with no identity switches.

It outperformed ByteTrack, StrongSORT, HybridSORT and DeepOCSORT when all methods ran on the same detections. The authors attribute the gains to the velocity-consistency fusion, which stabilizes association when apples are closely spaced, and to the lifecycle rule, which suppresses short-lived false tracks. Perception then had to become action.

The team introduced a trajectory-confirmation mechanism in which a track is declared defective once the cumulative number of defect-positive observations associated with it reaches a threshold K. Testing on production-line videos of 398 apples, including 335 defective and 63 normal samples, showed the trade-off clearly: with K equal to 1, no defective apple was missed but 36.51 percent of normal apples were falsely rejected; with K equal to 3, false rejections fell to 7.94 percent but 5.37 percent of defective fruit slipped through. K equal to 2 maintained zero missed rejections, reduced false rejections to 9.52 percent, and achieved the highest decision accuracy of 98.49 percent, so it was selected for the online test. The online validation exposed both the promise and the limits of the approach.

With the conveyor running at 0.16 meters per second and the paddle's open-close cycle measured at 0.6 seconds, corresponding to roughly 96 millimeters of conveyor travel per cycle, tightly spaced apples could reach the rejection position before the paddle fully returned. Of the 200 test apples, divided into four size classes of 50 each, the system's 15.0 percent missed rejection rate, driven by 12 missed defective fruit, far exceeded its 2.5 percent false rejection rate. The authors attribute this to the short distance between the paddle sorter and the fruit-cup unit, fruit spacing and pneumatic response delay.

The authors caution that validation covered a fixed imaging configuration and a limited operating range. Edge-of-image apples, where object scale shrinks further under long-range imaging, remain difficult for the detector. Broader testing across conveyor speeds, cultivars and operating conditions is still needed before the threshold strategy's operating boundary can be defined.

Within those limits, the prototype demonstrates that a single camera, a streamlined transformer and a simplified tracker can carry defect detection, tracking and timed ejection across three lanes in real time.

Subject of Research: Lightweight single-camera, three-lane machine vision and decision framework for online apple surface-defect sorting

Article Title: A lightweight perception and decision framework for single-camera three-lane online apple defect sorting

Article References: Qiao, P., Chen, G., Li, Z., Hao, J., Tang, W., Liu, Y., & Tchuenbou-Magaia, F. (2026). A lightweight perception and decision framework for single-camera three-lane online apple defect sorting. Artificial Intelligence in Agriculture. https://doi.org/10.1016/j.aiia.2026.09.009

Image Credits: AI Generated

DOI: 10.1016/j.aiia.2026.09.009

Keywords: apple sorting, machine vision, defect detection, object detection transformer, multi-object tracking, lightweight deep learning, agricultural robotics, post-harvest grading, RT-DETR, pneumatic sorting, precision agriculture, computer vision

A team at Jiangnan University’s School of Intelligent Manufacturing in Wuxi, China, has now proposed a way to soften that conflict at an unusually fine level of granularity. In a study published in Cluster Computing, Ziyi Zhao, Lei Su, and Ke Li introduce pFedDS, short for parameter domain sensitivity based personalized federated learning. Rather than treating a neural network as an indivisible bundle of weights to be averaged, their method interrogates each individual parameter, asking how strongly it reacts to the idiosyncrasies of a given client’s data, and then adjusts how that parameter is trained accordingly. The result, they report, is a measurable accuracy gain over eight state-of-the-art competing methods on two real industrial datasets.

To understand why this matters, it helps to consider how earlier attempts at personalization have worked. Many personalized federated learning schemes slice the model into two fixed territories: a shared portion, typically the lower layers that capture general features, and a personalized portion, often the upper layers or a dedicated head, which each client trains freely on its own data. The split is decided in advance and applied uniformly to every participant. The Jiangnan University authors argue that this static partitioning neglects the actual effects of data heterogeneity. A layer that is safely generic for one factory’s camera images may be deeply entangled with local quirks for another, and a one-size-fits-all boundary between shared and private knowledge cannot capture that variation.

The core innovation of pFedDS lies in how it measures sensitivity. During training, the method tracks each parameter’s deviation from the global update direction, the trajectory that the federated model as a whole would naturally follow. Parameters whose local updates stray far from that collective path are judged highly sensitive to the non-IID character of the client’s data. Those that hew closely to the global direction are deemed insensitive, meaning they encode knowledge that generalizes across participants. Crucially, this measurement is dynamic: it is recomputed as training proceeds, so the sensitivity profile of the model evolves alongside the data distribution and the optimization landscape rather than being frozen at initialization.

Once sensitivity has been quantified, pFedDS acts on it through learning rates. During local training, each parameter is updated with a sensitivity-aware step size. High-sensitivity parameters, the ones most entangled with local data peculiarities, receive larger learning rates, allowing them to drift toward local knowledge and specialize for the client at hand. Low-sensitivity parameters receive smaller learning rates, which restrains their movement and preserves the globally useful knowledge they carry. In effect, the method performs a continuous, per-parameter negotiation between global and local interests, replacing the blunt binary of shared versus personalized layers with a fine-grained spectrum in which every weight finds its own equilibrium point.

This design philosophy connects to a broader trend in the federated learning literature. The field has explored neuron-wise learning rates, parameter importance estimation borrowed from network pruning techniques, distribution-aware sub-model extraction, and adaptive local aggregation schemes, all in an effort to reconcile collective training with individual variation. What distinguishes pFedDS is the specific signal it uses, deviation from the global update direction, and the way that signal directly modulates optimization dynamics rather than merely deciding which parts of the model to share. The approach aims at adaptive knowledge fusion, letting global and local information blend in proportions that differ from parameter to parameter and from client to client.

The empirical case for the method rests on two practical industrial datasets rather than synthetic benchmarks. The first is a precision component defect dataset collected by the authors themselves, and the second is the Northeast University defect dataset, a surface defect collection widely used in industrial inspection research. Defect detection is a demanding testbed for federated learning because production lines differ enormously: lighting conditions, camera angles, material finishes, and the very types of flaws that appear vary from plant to plant, producing exactly the kind of severe data heterogeneity that breaks naive federated averaging.

Across these datasets, pFedDS delivered consistent improvements. Under pathological distribution settings, an extreme form of non-IID data in which each client holds samples from only a limited set of classes, the method achieved test accuracy between 1.34 and 1.4 percentage points higher than eight state-of-the-art baselines. Under extremely Dirichlet distributions, a statistically principled way of simulating heterogeneous client data, the advantage ranged from 0.89 to 1.12 percentage points. In a field where fractions of a percentage point can separate competing publications, gains of this size on two independent datasets, against a broad field of rivals, constitute a meaningful demonstration that sensitivity-aware personalization captures something that fixed partitioning schemes miss.

The industrial motivation behind the work is explicit. The authors are affiliated with a school of intelligent manufacturing, and the study was supported by the National Natural Science Foundation of China, the National Key Research and Development Program of China, the China Postdoctoral Science Foundation, and the 111 project, with acknowledgments extending to the Center for Advanced Life Cycle Engineering and the Centre for Advances in Reliability and Safety in Hong Kong. Factories are natural candidates for federated learning because production data is often commercially sensitive and cannot be freely pooled between companies or even between plants of the same company. A method that lets each site tune a shared inspection model to its own conditions, without shipping images anywhere, addresses both the privacy constraint and the heterogeneity problem at once.

Looking at the wider picture, pFedDS adds to a rapidly maturing toolkit for decentralized machine learning, a field whose open problems have been catalogued in a landmark survey by dozens of researchers and whose techniques now span adversarial domain generalization, knowledge distillation, meta-learning, and semi-supervised approaches. The persistent tension at the heart of the enterprise is that collaboration improves models while individuality degrades the shared objective, and every new method is essentially a different answer to how much of each to allow. By making that answer a continuously varying, data-driven property of every single weight, the Jiangnan University team offers one of the more granular answers yet proposed, and their results on real factory-floor data suggest the approach could matter wherever distributed devices must learn together while respecting the stubborn uniqueness of the data they hold.

Subject of Research: Parameter domain sensitivity based personalized federated learning for non-IID data

Article Title: pFedDS: parameter domain sensitivity based personalized federated learning

Article References: Zhao, Z., Su, L., & Li, K. (2026). pFedDS: parameter domain sensitivity based personalized federated learning. Cluster Computing, 29(13), Article 761. https://doi.org/10.1007/s10586-026-06578-9

Image Credits: AI Generated

DOI: 10.1007/s10586-026-06578-9

Keywords: federated learning, personalized federated learning, non-IID data, parameter sensitivity, machine learning, defect detection, industrial inspection, distributed training, data privacy, adaptive learning rates, Cluster Computing, pFedDS

The research, led by Kao Ge and Qing-Bang Han, tackles a problem that has long frustrated engineers: sonar is the only practical sensor in fully flooded, opaque pipes, but sonar images are notoriously ugly to look at and harder to interpret. Acoustic imaging produces grainy, low-contrast pictures in which a small fracture can occupy just a handful of pixels against a background of acoustic speckle, reverberation from pipe walls, and noise from the water itself. Conventional computer vision models, trained mostly on crisp optical photographs, tend to collapse under these conditions. Deep learning models that do work well on sonar are usually too large and power-hungry to embed in the battery-constrained vehicles that would actually crawl through a municipal drain.

The Hohai team’s answer is an architecture they call ShuffleNet-MSAA, and its design reflects a careful trade-off between intelligence and efficiency. The backbone of the network is ShuffleNet, a convolutional neural network originally engineered for smartphones and other mobile devices. ShuffleNet achieves its speed through a clever trick known as channel shuffling, in which information is mixed across feature channels in a way that preserves accuracy while drastically cutting the number of computations. That makes it an ideal starting point for a detector that must live on embedded hardware rather than in a data center. But a lightweight backbone alone cannot solve the fundamental signal problems of sonar imagery, so the researchers layered two specialized modules on top of it.

The first is a multi-scale feature extractor. Defects in a drainage pipe come in wildly different sizes: a hairline crack may span a few pixels while a collapsed section dominates the frame. Networks that look at an image through a single receptive field tend to miss one extreme or the other. By extracting and fusing features at multiple scales simultaneously, the new model can attend to both tiny anomalies and large structural failures within the same pass. The authors report that this multi-scale strategy was particularly important for small targets, which are precisely the defects most likely to be overlooked by human inspectors and by earlier automated systems alike.

The second innovation is the multi-scale adaptive attention module, the MSAA of the system’s name. Attention mechanisms in modern AI work roughly the way human attention does: they let a network decide which parts of an input deserve focus and which can be ignored. What distinguishes the new module is that it operates pixel by pixel and adapts dynamically to each image, weighting the contribution of every location according to how likely it is to contain genuine defect information rather than noise. Notably, the researchers implemented the module using median filtering, a classical signal-processing technique prized for suppressing impulsive noise without blurring edges. Combining this statistical robustness with learned attention allows the network to quiet the acoustic clutter of a turbid pipe while sharpening the faint signatures of real defects.

Even a well-designed network can be undermined by the data it learns from. Sonar defect datasets are inherently imbalanced: catastrophic failures are rare, minor defects are common, and some defect categories may be represented by only a handful of examples, producing what statisticians call a long-tailed distribution. Standard training procedures end up favoring the frequent categories and neglecting the rare ones, which is exactly backwards for infrastructure monitoring, where catching an uncommon but severe fault matters most. To counter this, the team introduced a dual-weighted focal loss, or DW-FL, a modified training objective that extends the focal loss concept originally developed for dense object detection. The dual weighting scheme adjusts the loss both for class frequency and for example difficulty, forcing the network to invest its learning capacity in the rare, hard-to-classify defects that conventional training would gloss over.

The experimental results are the strongest part of the story. The researchers evaluated ShuffleNet-MSAA on both a custom dataset of drainage pipe sonar images and a public sonar defect benchmark, comparing it against state-of-the-art baselines built on both convolutional neural networks and Transformer architectures. The proposed system outperformed these competitors in both detection and classification performance, while remaining lightweight enough for deployment on resource-constrained platforms. The authors emphasize that the model showed high accuracy, robustness, and generalization across the two datasets, a combination that matters because a detector tuned to one pipe network’s acoustics often fails when moved to another. The work was supported by the Natural Science Foundation of China, the Key Research and Development Project of Changzhou in Jiangsu Province, a Jiangsu Provincial graduate innovation program, and the company AutoSubsea Vehicles Inc., an affiliation that hints at commercial ambitions for the technology.

What makes this study worth wider attention is how it fits into a broader shift in how societies maintain their buried infrastructure. Cities worldwide are grappling with aging water systems, and the cost of undetected pipe failures, from sinkholes to sewage contamination, runs into billions annually. Earlier automation efforts relied on closed-circuit television, which works only in dry or near-dry pipes, or on hand-crafted acoustic features that required expert tuning. The new work builds on a decade of progress in sonar deep learning, including transfer learning approaches, hybrid CNN-Transformer frameworks, and attention modules adapted from general computer vision, but it is among the first to target the specific, punishing conditions of underground drainage: full submersion, heavy sediment, severe noise, and a hard ceiling on compute. The pixel-wise adaptive attention design also echoes a trend in medical imaging, where adaptive spatial weighting has recently improved segmentation under noisy conditions, suggesting a convergence of techniques across domains that share the same core problem, extracting weak signals from hostile data.

There are, of course, caveats. The published results come from curated datasets, and the true test will come when the system meets the chaotic reality of a working sewer, with its shifting debris, variable flow, and acoustic surprises no benchmark anticipated. The authors state that their data are available upon request, which should allow independent groups to probe the model’s limits. Still, the engineering philosophy on display, pairing an efficient mobile-grade backbone with task-specific attention and an imbalance-aware loss, offers a template that other underwater robotics applications, from subsea pipeline leak detection to seabed mapping, could readily adopt. As autonomous inspection vehicles grow cheaper and acoustic sensors grow better, the bottleneck is increasingly the intelligence onboard. Work like ShuffleNet-MSAA suggests that bottleneck is now being cleared, one carefully weighted pixel at a time, and that the hidden arteries beneath our cities may soon be examined routinely, safely, and without a single human entering the water.

Subject of Research: Lightweight deep learning for sonar-based defect detection in underground drainage pipelines

Article Title: ShuffleNet-MSAA: a lightweight multi-scale adaptive attention mechanism for enhanced sonar-based defect detection in underground drainage pipelines

Article References: Ge, K., & Han, Q.-B. (2026). ShuffleNet-MSAA: a lightweight multi-scale adaptive attention mechanism for enhanced sonar-based defect detection in underground drainage pipelines. Multimedia Tools and Applications, 85(9), Article 739. https://doi.org/10.1007/s11042-026-21816-3

Image Credits: AI Generated

DOI: 10.1007/s11042-026-21816-3

Keywords: sonar, defect detection, drainage pipelines, ShuffleNet, attention mechanism, deep learning, multi-scale feature extraction, focal loss, underwater acoustics, pipeline inspection, lightweight neural networks, computer vision