Abstract:Hydraulic machinery and equipment (such as gates, trash racks, etc.) inevitably suffered from corrosion on their structural surfaces due to the long-term usage in the water environment. If the severity of corrosion cannot be accurately detected, the maintenance and strengthening of important key stress-bearing structural components may fail to be timely implemented, which further threatens people’s life around the damaged structure. Currently, the identification of corrosion features of hydraulic machinery and equipment mainly depends on manual visual inspection, resulting in strong subjectivity, poor detection accuracy limited by visual fatigue. To address this issue, a corrosion-rating evaluation method is proposed by incorporating the attention mechanism and bilinear pooling into the VGG-16 (visual geometry group, VGG) backbone. Considering that different components in the RGB (red green and blue) and HSV (hue saturation and value) color spaces present distinctive features on the rust image, a multi-branch neural architecture is designed to learn informative features by taking different color spaces as inputs. The attention mechanism is embedded in the branch network to recalibrate the features of the rust image by the learnable weights, focusing on the learning of task-oriented features to support the evaluation. The bilinear pooling is applied to fuse the extracted representations of branch networks to enhance the fine-grained features, which improves the ability to utilize the contrastive details. A rust image dataset is constructed to validate the proposed approach using the salt spray corrosion test . Experimental results demonstrate that the proposed approach outperforms the selective baselines, achieving a 0.953 detection and other measurements, including the precision, recall and F1 scores, are also greatly improved. Most importantly, the proposed approach can achieve better performance for irregular corrosions, which has huge potential for the application of the engineering practice.