RTDETR-Bolt: a lightweight and efficient method for bolt surface defect detection

Abstract To address challenges in vision-based bolt surface defect detection—including inaccurate feature extraction due to irregular defect shapes and multi-scale features, interference in defect localization caused by complex backgrounds, and reduced detection accuracy resulting from lightweight models—this paper proposes a lightweight and efficient defect detection algorithm (RTDETR-Bolt) based on the RT-DETR model. This approach significantly reduces model complexity while maintaining detection accuracy, demonstrating strong potential for edge deployment. First, a channel pruning strategy is introduced to ResNet-18, designing the lightweight convolutional module partial convolution Block. Combined with the global and local feature merging module, this forms the efficient backbone network LightPGNet, which enhances feature extraction capabilities while substantially reducing model complexity. Second, we propose the dynamic scale sequence feature fusion (SSFF) structure, dynamic group convolution fusion module, and collaborative channel-spatial-pixel attention mechanism. These are integrated into the dynamic dimensional collaboration-based scale sequence feature fusion (DDC-SSFF) architecture, enabling efficient fusion of multi-scale information and precise capture of defect features. Finally, the Distance Intersection over Union loss function is introduced to enhance target localization accuracy, while transfer learning strategies accelerate model convergence and mitigate performance degradation caused by lightweight optimization. Experimental results demonstrate that RTDETR-Bolt achieves outstanding performance in bolt surface defect detection. Compared to baseline algorithms, the proposed approach reduces model parameters and computational complexity by 72.0% and 48.5%, respectively, while boosting detection speed by 1.2 times. It achieves a detection accuracy of 95.3%, representing a 0.3% improvement over the original model, thereby maintaining high-precision detection capabilities while achieving lightweight optimization.