LiMS-Net: Lightweight metal surface defect detection network

Abstract

This study aims to enhance the accuracy of detecting defects on metal surfaces by proposing a lightweight metal surface defect detection network (LiMS-Net). The backbone of LiMS-Net incorporates a residual synchronous convolutional block feature extraction module that utilizes multi-scale convolution kernels. Features are concurrently processed using these multi-scale convolution kernels. In the neck stage, a Conv-MLP module that extracts global image features. This module is further enhanced by shift operations that improve information interaction among different regions of the features. To further enhance feature interaction across different scales and improve detection accuracy, a cross-scale feature fusion block is proposed. This approach alleviates feature loss issues caused by extensive feature processing. This study employed the advanced object detection methods and conducted comparative experiments using publicly available defect databases. Compared to the advanced object detection methods, LiMS-Net demonstrated superior performance across all databases while utilizing fewer parameters.