A Lightweight Residual Model for Corrosion Segmentation with Local Contextual Information

Metal corrosion in high-risk areas, such as high-altitude cables and chemical factories, is very complex and inaccessible to people, which can be a hazard and compromise people’s safety. Embedding deep learning models into edge computing devices is urgently needed to conduct corrosion inspections. However, the parameters of current state-of-the-art models are too large to meet the computation and storage requirements of mobile devices, while lightweight models perform poorly in complex corrosion environments. To address these issues, a lightweight residual deep-learning model based on an encoder–decoder structure is proposed in this paper. We designed small and large kernels to extract local detailed information and capture distant dependencies at all stages of the encoder. A sequential operation consisting of a channel split, depthwise separable convolution, and channel shuffling were implemented to reduce the size of the model. We proposed a simple, efficient decoder structure by fusing multi-scale features to augment feature representation. In extensive experiments, our proposed model, with only 2.41 MB of parameters, demonstrated superior performance over state-of-the-art segmentation methods: 75.64% mean intersection over union (IoU), 86.07% mean pixel accuracy and a 0.838 F1-score. Moreover, a larger version was designed by increasing the number of output channels, and model accuracy improved further: 79.06% mean IoU, 88.07% mean pixel accuracy, and 0.891 F1-score. The size of the model remained competitive at 8.25 MB. Comparison work with other networks and visualized results were used for validation and to determine the accuracy of metal corrosion surface segmentation with limited resources.