Optimized YOLO and Semi-Supervised Fuzzy Graph Convolutional Network (SSFGCN) For Surface Defect Recognition

Abstract

In the manufacturing sector, surface defect detection is a vital technique for the integrity of goods. The contemporary manufacturing sector demands great examination accuracy; conventional manual visual inspection cannot match these standards. Moreover, to satisfy the demands of strict quality assurance criteria, excellent quality input images from excellent quality industry cameras are required. Therefore, it is crucial to create an accurate object sensor that can take in excellent-quality information and infer it in real time. The primary aim of this work is to distinguish identical goods, continuous motion, and poor excellence. Optimized You Only Look Once (OYOLO), YOLOv3 anchor box size is optimized using Dove Swarm Optimization (DSO). The proposed framework included an improved anchor box size, a multiple-scale sensing design, and a compact backbone built on a sparsely linked neural network. The Next Zero-Shot challenge in surface defect detection utilizes the Optimized You Only Look Once Zero-Shot with Class Knowledge Graph (OYZS-CKG). OYZS-CKG will establish a category knowledge graph and determine the connection between novel defect classes and base. Afterward, a Semi-supervised Fuzzy Graph Convolutional Network (SSFGCN) is introduced to extract features of classes. Extensive tests are conducted to illustrate the effectiveness of the suggested strategy on two datasets: NEU-CLS and Xsteel Surface Defect (X-SSD). The performance evaluation metrics are precision, recall, F1-score, and accuracy to assess its impact appropriately. The proposed approach achieved the highest accuracy compared with other state-of-the-art approaches: SGCN, Graph Convolutional Network Zero-shot (GCNZ), and Zero-Shot with Class Knowledge Graph (ZS-CKG).