Modified Vision Graph Neural Network-based Steel Surface Defect Classification

The steel defect is a common condition in the smelting and forging process that causes non-negligible loss for the manufacturer. Computer vision-based defective steel identification methods provide efficient and reliable solutions for detecting steel quality, mainly including traditional machine vision and deep learning. However, most of these methods are limited by the feature engineering and over-fitting problems, the balance between recognition accuracy and computational cost cannot be well achieved. Therefore, it is necessary to develop a method that not only efficiently steel defect identification, but also overcomes the heavy memory and computational requirements. In this paper, we propose a modified vision graph neural network that can efficient and flexibly captures irregular and complex image information for classifying defective steel and normal steel and not cost excessive computing resources. Compared with other works, our improvement lies in three aspects: 1) apply vision graph neural network to capture inner relation of irregular defect; 2) ameliorate Stem stage to providing better feature representations; 3) integrated use max and mean pooling to aggregate rich neighbor node information. Experiments demonstrate that our improved method is not increase excessive calculation demands and efficient in results on the recognition of steel surface images.