A novel multi-scale cross-patch attention with dilated convolution (MCPAD-UNET) for metallic surface defect detection

Abstract

Surface defect detection in industrial processes is crucial for ensuring product quality and reducing material waste. Automated defect identification using deep learning techniques has become a vital aspect of the automated surface defect detection field. However, achieving accurate and automatic defect segmentation remains a significant challenge, especially for fine precision segmentation required in high-quality products. The traditional approaches for defect segmentation have several limitations, such as difficulty in preserving fine details and contextual information, leading to poor segmentation performance. To overcome these limitations, new segmentation algorithms that can preserve fine precision and contextual information need to be evaluated. Therefore, there is a need for novel segmentation algorithms that can accurately identify and segment defects in industrial processes, incorporating multi-scale contextual information, preserving fine details, and handling complex and subtle defects. In this paper, we propose a novel approach for steel defect segmentation called multi-scale cross-patch attention with dilated convolution (MCPAD-UNet). This approach employs a subsampled module that achieves the same dimensionality reduction as max-pooling while preserving the fine precision of the features. Additionally, MCPAD-UNet utilizes a cross-patch attention module with dilated convolution, simultaneously collecting channel–spatial data and integrating relevant multi-scale features to reduce the semantic gap and enhance detailed information. To prevent overfitting, we apply dropout after each hybrid dilated convolution block. Extensive testing on the public Severstal: Steel Defect Detection dataset demonstrates the effectiveness of our approach, achieving Dice scores of 95.3%, outperforming the competition's overall score by 5.2%. Our proposed method has the potential to significantly improve defect detection in industrial processes, thereby reducing material waste and improving product quality. © 2023, The Author(s), under exclusive licence to Springer-Verlag London Ltd., part of Springer Nature.