Angle Steel Tower Bolt Defect Detection Based on Transformer

Abstract

The bolts of the angle steel tower will rust, loose, and fall off under natural conditions. Traditional manual bolt defect detection is inefficient and dangerous. This paper proposes CViT-FRCNN based on ViT-FRCNN, which uses a convolutional neural network as the backbone model and the output features are input to the Transformer encoder. Compared with the direct patch embedding of ViT-FRCNN, this can improve the richness of input features and detection accuracy. A series of experiments show that our proposed model achieves the best performance in angle steel tower bolt defect detection and meets the needs of power inspection scenarios.