Edge devices-oriented surface defect segmentation by GhostNet Fusion Block and Global Auxiliary Layer

This paper introduces an approach for the segmentation of surface defects, referred to as Efficient Surface Defect Network (ESD-Net). The proposed method uses novel components called the GhostNet Fusion Block (GFB) and the Global Auxiliary Layer (GAL) to make it edge computing-ready and to increase its performance on segmentation. The GFB algorithm employs a technique whereby it conserves and combines feature maps of reduced resolution from the original image with feature maps that have been downsampled at various resolutions. Moreover, the GAL amplifies the GFB by including comprehensive contextual information from a global perspective. The experiment shows that the proposed method outperforms state-of-the-art algorithms on SD-saliency- 900, MSD, and Magnetic-tile, three public surface defect datasets with mIoU of 82.4%, 92.9%, and 78.8%, respectively. Embedded device experiments have proven that ESDNet can be utilized on a wider range of cost-effective industrial devices with acceptable latency.