An efficient multiscale enhancement network with attention mechanism for aluminium defect detection

Real-time defect detection is required to efficiently control the quality of aluminium. However, aluminium defects have the characteristics of small-size, low contrast, and multiscale variations, which pose great challenges to defect detection. This article aims to improve the defect detection accuracy and propose an effective multiscale enhancement network with attention mechanism for aluminium defect detection (AMMENet). First, to capture key features and mitigate interference from the background, a pluggable parallel residual attention module (PRAM) is proposed for the feature extraction network. To compensate for the loss of deep features, a multilevel semantic enhancement module (C2f-MFF) is proposed to fuse multiscale feature maps. Finally, the model was applied to the Tianchi Aluminium Surface Defect Dataset (TC-ASDD) for ablation experiments and comparisons. The experimental results show that the mean average precision (mAP@0.5) of the proposed AMMENet is 73.6% with a real-time detection speed of 66.2 frames per second (FPS). Compared with YOLOv8 baseline network, AMMENet improves mAP@0.5 by 2.8% with only a slight loss in speed. Moreover, AMMENet is superior to the state-of-the-art detection methods in terms of detection accuracy.