The missing of nuts and bolts is a common structural defect in the construction stage of angle steel towers, but the detection rate of bolt defects by existing object detection algorithms is low due to low feature discrimination. In order to solve this problem, a global information extraction operator is proposed based on Transformer to encode convolutional features. Secondly, the local background information introduced after the multi-scale scaling of the candidate detection frame is adaptively combined through the channel attention mechanism. Finally, the bolt defect samples are amplified based on image segmentation and background fusion. The ablation experiments show that the above strategies can effectively improve the detection effect of bolt defects and do not exclude each other. Compared with other typical algorithms, this algorithm has been proven to be advanced.