Research on Machine Vision-Based Disordered Steel Detection Using an Improved PatchCore Method

In steel bar production, disordered steel on the cooling bed can reduce production efficiency and cause economic losses. Current detection methods rely mainly on manual visual inspection, which suffers from low timeliness and accuracy due to fatigue in high-temperature, high-dust environments. This makes it difficult to meet the demands of machine vision-driven intelligent quality inspection. To address this issue, we propose an improved PatchCore-based method for disordered steel detection. The core process of this method includes the following steps: First, inverse perspective transformation is used to correct the perspective distortion of images captured by cameras, reprojecting the images to a top-down view of the cooling bed. Then, the images are divided into local image patches using an n \(\times \) n grid to increase the proportion of abnormal regions and enrich the training samples. Second, features of the local image patches are extracted, and a selected memory bank is constructed. K-means clustering is used to replace the original coreset sampling; after preprocessing steps such as dimensionality reduction and standardization of the local image patch features, clustering is performed to generate the selected memory bank, balancing feature coverage and inference efficiency. Finally, data preprocessing, feature extraction, and anomaly determination are conducted on the images to be detected. In the inference stage, the Euclidean distance between the features of the test image patches and the features in the memory bank is calculated as the anomaly score, and a statistical distribution-based threshold strategy is used to determine whether an image is abnormal. Image partitioning increases the proportion of abnormal region area in abnormal images from 0.26% to 1.69%. The results show that the proposed model achieves an AUPR of 0.997, an AUROC of 0.996, and an FPS of 4.397, all outperforming the comparison models. The improved PatchCore method proposed in this paper effectively addresses the limitations of traditional manual detection and the adaptability issues of existing models, significantly improving the accuracy of disordered steel detection and the feasibility of practical deployment, thus providing technical support for the stable operation of the cooling bed system.