Improved YOLOv5 Surface Defect Detection Algorithm for Aluminum

Abstract

Aluminum is widely used in industry, and for the current problems of low accuracy and poor generalization ability of the traditional method to detect surface defects in aluminum. In this paper, we propose an improved YOLOv5s model for detecting surface defects on aluminum. Firstly, the RFB module (Receptive Field Block) is introduced in the backbone network, and the RFB is appropriately adjusted to bring a larger perceptual field according to the characteristics of the scale of aluminum surface defects to enhance the feature discriminability. Secondly, the channel attention mechanism (Squeeze-and-Excitation) is introduced into the C3 module, and the C3SE module is designed to make the network pay more attention to the features with high weights. Finally, the coupled detection head is replaced with a more efficient and streamlined decoupled head, which can better adjust the position of the target frame precisely and improve the detection accuracy. The experimental results prove that the improved model performs well, with [email protected] reaching 81%, which is 3.2% better than the original model, while ensuring the speed of detection and meeting the requirements of aluminum detection.