Deep Learning Model Fused with Attention Mechanism for Defect Detection of Electroplated NdFeB Products

Abstract

Electroplated neodymium-iron-boron (NdFeB) products are widely used in electronics, automobiles, and new energy industries. Surface defects, such as cracks, pinholes, scratches, and edge chipping, can directly affect product performance and service life. To address the problems of insufficient detection accuracy, weak anti-interference capability, and difficulty in identifying tiny defects in traditional deep learning models for NdFeB surface defect detection, an improved defect detection model integrating an attention mechanism is proposed. Based on YOLOv8n, the proposed model introduces an improved Recurrent Channel Attention (RCA) mechanism into the backbone network to enhance the extraction of critical defect features and suppress complex background interference. Meanwhile, Switchable Atrous Convolution (SAConv) is incorporated to enlarge the receptive field and improve the model's perception and representation ability for multi-scale defects. Experimental results show that the proposed model achieves an mAP@0.5 of 85.3% on a self-built electroplated NdFeB defect dataset, representing an improvement of 6.7% over the original YOLOv8n model. The detection speed reaches 62 FPS, meeting the real-time detection requirements of industrial production lines. The proposed model can effectively identify various common surface defects in electroplated NdFeB products and demonstrates strong detection capability for tiny and subtle defects under complex lighting conditions, thereby providing reliable technical support for intelligent quality inspection of electroplated NdFeB products.