Segmentation and grade evaluation of corrosion on hydraulic steel gates based on image-level labels

Machine vision offers distinct advantages, such as enhanced efficiency and precision, in the segmentation and assessment of corrosion on hydraulic steel gates. This study addresses the challenge of demanding a substantial amount of pixel-level annotated data in machine vision-based corrosion segmentation and assessment approaches. To tackle this issue, a novel weakly supervised method for corrosion segmentation and assessment in hydraulic steel gates is proposed, leveraging class labeling. The technique employs a class activation map to pinpoint regions containing corrosion seeds and to train a network to capture semantic affinity relations. Subsequently, the concept of region growing is adopted to propagate semantic information across the entire image. The average feature vector of the seed region serves as the corrosion feature, enabling precise segmentation of corroded areas and circumventing the laborious pixel-level annotation process. Additionally, a fine-grained corrosion classification network is established and trained using salt spray corrosion test data to accurately evaluate the corrosion severity. To validate the proposed method's accuracy, a dataset of steel gate corrosion images is curated based on real-world operational scenes. Experimental results demonstrate that, in practical scenarios, the segmentation method presented in this paper achieves a segmentation intersection ratio of 62.37% in corrosion, without pixel-level annotation. This performance closely approaches the performance of mainstream fully supervised methods. Additionally, the corrosion grade evaluation method proposed in this study achieves an accuracy of 95.77%.