The Appearance Defect Detection Recognition and Classification of Fuses Based on Deep Learning

This study is to extract image data of fuses to train an improved object detection model, to enhance the efficiency and accuracy of appearance defect detection in fuse production lines. A fuse appearance defect classification model is developed using deep learning techniques. Improvements are made to the original YOLOv8 model by incorporating a Multi-space Pyramid Pooling Structure (M-SPP) and an attention module. The M-SPP structure is designed to strongly capture multi-scale features, at the same time the attention module is introduced to direct the model's focus more effectively on target regions. Additionally, the Conditional Generative Adversarial Network (cGAN) algorithm is employed for data augmentation, and modifications are made to the Non-Maximum Suppression (NMS) algorithm to minimize redundant detection boxes and improve detection accuracy. Experimental results demonstrate that the modified target detection model achieves an average accuracy of 99.68% and can handle one potential faulty piece per 0.5 seconds. This proposed model with high accuracy and rapid detection speed is well-suited for real-time automated detection in fuse production lines.