The research team, led by Kesheng Wang of Shandong Huayu University of Technology together with colleagues at the North China Institute of Aerospace Engineering and Space Engineering University, built GSDD-Net on top of the YOLO family of one-stage object detectors, which predict object classes and bounding boxes directly from feature maps rather than working through a slower region-proposal stage. That choice matters for factories, where inspection systems must keep pace with production lines in real time. But off-the-shelf detectors struggle with gear defects, which vary enormously in scale and often occupy only a handful of pixels in an image. The team’s answer was not one big idea but four targeted architectural changes, each aimed at a specific failure mode of existing models.

The first change targets a deceptively mundane step: downsampling. Neural networks shrink images as they process them, and conventional strided convolutions throw away fine detail along the way, exactly the edge contours and texture cues that distinguish a faint scratch from ordinary machining marks. GSDD-Net replaces standard downsampling with a module called ADown, which splits the input features into two branches, sends one through convolution and the other through max pooling, and concatenates the results. By combining average-style global context with locally salient responses, ADown preserves weakly salient defect information while cutting computational cost. In the team’s ablation tests on the GID dataset, ADown alone reduced the parameter count from 2.6 million to 2.1 million while still nudging accuracy upward.

The second component, CSP-EDLAN, reworks the feature extraction and fusion stages in both the backbone and the neck of the network. It folds together three ideas: cross-stage partial connections that give gradients an additional route to flow during training, multi-gradient path aggregation that merges features from different branches, and DualConv units that pair a grouped spatial convolution with a pointwise convolution so channels can exchange information cheaply. The result is richer multi-scale feature representations for defects of diverse sizes and shapes, with less redundant computation. Among all four modules tested individually, CSP-EDLAN delivered the largest accuracy gain, lifting the key mAP50 metric by 1.5 percentage points over the YOLO11n baseline.

The third innovation confronts the messy optics of real factory floors. Metallic reflections, grease, and non-uniform illumination generate spurious responses that can swamp the weak signal of a small defect. To handle this, the researchers swapped the attention block in the baseline network for a new module called C2TSSA, built on token statistical self-attention. Instead of computing attention by comparing every pair of image tokens, an operation whose cost grows quadratically, C2TSSA derives attention weights from token-level statistics such as the mean and covariance of the features, capturing global context with only linear computational growth. The module strengthens responses consistent with the surrounding gear structure while suppressing isolated responses caused by reflections and background textures, making subtle defects stand out against their surroundings.

The fourth component, LQEHead, tackles a mismatch that plagues many modern detectors: the classification branch and the box-regression branch work separately, so a high confidence score does not always mean the predicted box actually sits on the defect. This matters acutely for small flaws, where even a slight box displacement sharply reduces overlap with the true region. LQEHead extracts stability statistics from the predicted bounding-box distributions and converts them into a signed correction that is added to the classification logit before the final activation. Predictions with reliable localization get a confidence boost; unstable ones are dialed down. The design preserves weak but well-placed detections while filtering out spurious responses from background edges and glare.

Put together, these four modules produced striking results on two benchmarks. On the GID dataset, a set of 2,000 industrial gear images covering chipping and two types of black-skin defects, GSDD-Net achieved an mAP50 of 73.0 percent and an mAP50-95 of 38.0 percent, beating the YOLO11n baseline by 3.1 and 2.1 percentage points respectively. Meanwhile the model shrank: parameters fell from 2.6 million to 1.9 million, computational cost dropped from 6.3 to 5.9 GFLOPs, and inference speed climbed from 93.7 to 105.7 frames per second. On a second, publicly available gear defect dataset featuring break, missing-tooth, and scratch defects, GSDD-Net reached 97.8 percent mAP50 and 69.6 percent mAP50-95, again improving on the baseline while using fewer parameters.

Comparisons against a broad field of contemporary detectors underscore the balance the team struck. Larger and heavier models such as RT-DETR-l, which carries 32 million parameters and needs more than 100 GFLOPs, failed to match GSDD-Net’s accuracy while running far slower at 38.8 frames per second. Newer YOLO variants including YOLOv12n, YOLOv13n, and YOLO26n each made different trade-offs between speed and accuracy but none matched GSDD-Net’s combined mAP scores on the GID dataset at comparable cost. Class-by-class analysis showed consistent gains: the hardest category, chipping, improved by 4.3 percentage points in AP50, and the stricter AP50-95 metric rose for every defect class, indicating better localization, not just better recognition. Grad-CAM heatmaps provided visual confirmation, showing GSDD-Net concentrating its attention squarely on defect regions while the baseline scattered its responses across irrelevant background.

The authors are careful to frame these numbers as results under their evaluated conditions, noting that deep-learning outcomes can fluctuate with random initialization, data augmentation, and GPU nondeterminism, and that all compared models were trained under identical settings to ensure fairness. They also acknowledge limitations: the datasets cover a limited set of defect categories and imaging conditions, and real factories may present harsher variation. Future work, they write, should expand datasets, explore semi-supervised and synthetic data generation to ease the burden of annotation, apply pruning and quantization for deployment on edge hardware, and even integrate the detector with large language models to build an intelligent inspection agent.

Still, the broader significance is clear. Quality control in intelligent manufacturing has been squeezed between two bad options: accurate but slow and power-hungry models, or fast but less reliable ones. By showing that task-oriented design, a pooling scheme that preserves detail, an aggregation module that enriches features, a statistical attention mechanism that reads the whole scene cheaply, and a detection head that knows how sure it should be, GSDD-Net demonstrates that precision and efficiency need not be enemies. For production lines where a missed scratch on a gear tooth today can mean a seized gearbox tomorrow, that balance could translate directly into fewer failures, less downtime, and safer machines.

Subject of Research: Lightweight deep learning framework for detecting gear surface defects in industrial images

Article Title: GSDD-net: an efficient feature extraction method for gear-surface defects in complex industrial environments

Article References: Wang, K., Zhang, H., Yang, T., Shen, P., & Zhao, Y. (2026). GSDD-net: an efficient feature extraction method for gear-surface defects in complex industrial environments. Frontiers in Mechanical Engineering, 12, Article 1983450. https://doi.org/10.3389/fmech.2026.1983450

Image Credits: AI Generated

DOI: 10.3389/fmech.2026.1983450

Keywords: GSDD-Net, gear defect detection, YOLO, deep learning, machine vision, industrial inspection, object detection, self-attention, lightweight models, predictive maintenance, feature extraction, surface defects