Aiming at the problem of low recognition rate of traditional target detection algorithm for inner wire joint sealing surface defects, an improved YOLOv4 algorithm was proposed to detect the defects. Firstly, k-means++ clustering algorithm is used to optimize the parameters of the anchor frame of the target sample, and improve the matching degree between the anchor frame and the feature map; Secondly, the SENet attention mechanism module is introduced into the backbone network to strengthen the key information of the image, suppress the background information of the image, and improve the confidence of the defect that is not easy to identify; after that, the SPP module is added to the neck of the network to enhance the acceptance domain of the backbone network output features and separate the important context information; Finally, using the collected data set of inner wire joint sealing surface defects to train the original YOLOv4 and the improved YOLOv4, and the performance of models were tested respectively on test set. The experimental results show that the performance of YOLOv4 is good, but some small targets are missed; The improved model has excellent detection performance for small target defects, the mean average accuracy (mAP) reaches 87. 47%, which is 10. 2% higher than the original YOLOv4, and the average detection time is 0. 132 s, which realize the rapid and accurate detection of inner wire joint sealing surface defects.