NBD- YOLOv5: An Efficient and Accurate Aluminum Surface Defect Detection Method

To address the challenges associated with low accuracy in detecting surface defects on aluminum materials, particularly the tendency to overlook small target defects, we present an improved defect detection method called NBD-YOLOv5. Our approach incorporates several enhancements to the original YOLOv5s model. Firstly, we tackle issues such as gradient vanishing and feature redundancy by replacing the backbone feature extraction network with DenseNet. This modification significantly improves the overall performance of the model. To boost the model's learning capability for small target defects on aluminum surfaces, we introduce the Normalized Wasserstein Distance (NWD) metric into the regression loss function. This addition enhances the model's sensitivity to subtle defects. Moreover, we integrate the latest Biformer dynamic attention mechanism into the network. This mechanism adjusts weights adaptively based on image features, thereby improving the model's detection capabilities. In an effort to enhance convergence speed and accuracy during the training process, we decouple the head of YOLOv5. This adjustment further refines the model's performance. Finally, we validate our approach using the aluminum material classification dataset from the preliminary round of the Guangdong Industrial Smart Manufacturing Big Data Innovation Competition. Our method achieves an 8.1% increase in mean Average Precision (mAP) compared to the original YOLOv5 model, demonstrating clear advantages over existing mainstream detection methods.