<p>Due to the low recognition rate of traditional object detection algorithms for target defects, an improved Faster-RCNN algorithm is proposed for detecting defects on the sealing surface of inner wire joints. SENet attention modules are embedded in the backbone of Faster-RCNN to enhance the confidence of objects that are difficult to recognize by enhancing key image information and suppressing background information. Then initial feature maps of multiple different scales from the backbone network are bidirectionally fused by an improved FPN network to obtain the final feature map containing strong position information and strong semantic information to improve the utilization of small object features. ROI Align is used to eliminate the quantization errors introduced during the pooling process of regions of interest. Model training is improved by redesigning the smooth L1 loss function to balance it for optimization that would effectively reduce the imbalance between difficult-to-learn samples with a large gradient and easy-to-learn samples with a small gradient. Based on the dataset constructed, the ablation experimental analysis verified that the proposed method effectively improves the automatic detection of defects in the surfaces of inner wire joints. Specifically, mAP increases by 11.85% compared to that of Faster-RCNN. Compared with other methods, the proposed method provides better performance in detecting accuracy.</p>