Corrosion Identification of Fittings Based on Computer Vision

The metal rust environment is complex, and the rust parts and shapes are quite different, making rust difficult to detect. As drones are gradually applied to line inspections, computer vision can be used for the identification of metal corrosion. Aiming at the problems existing in the current corrosion detection, this paper proposes a corrosion detection algorithm based on Faster-RCNN target detection model and the rust HSI color feature, which is used to solve the problem of poor applicability and inefficiency of digital image processing and features cannot be accurately extracted when using deep learning method and other issues. First, the rust image is converted from the RGB color model to the HSI color model, and then each pixel of the HSI space is traversed. According to the threshold range of the rusted color feature, it is determined whether the pixel is rusted, thereby removing the complex interference background in the image, leaving only the rusted area used to facilitate the labeling. Then, manually labeling into a training set through the LabelImg open source annotation tool. The labeled data set facilitates feature extraction by convolutional neural networks because only rusted areas are present. Corrosion detection and localization were performed on the prepared training set using the Faster-RCNN target detection model. The results show that it has a good recognition effect for several common rust conditions. Moreover, the method of combining the deep learning algorithm with the HSI color feature achieves a high level in determining the correctness and recall rate of rust, and the leak recognition rate also meets the practical requirements.