Abstract: The single view detection based on convolutional neural network cannot effectively identify 3D shape defect targets. As a result, in practical applications, the low-precision detection requirements are often achieved by detecting only one of the most representative views or detecting each face in turn, which brings large time cost and use restrictions. To solve this problem, this paper proposes an improved residual network and applies it to 3D shape solder joint defect detection. The model can firstly obtain all view images of solder joints at one time, and then get detection results through feature aggregation and adaptive learning module. The multi-view solder joint data set was obtained from high frequency inductor element, and the identification accuracy of the proposed method reached 99.48%. The results show that the improved residual network can effectively improve the image identification accuracy under the same number of network layers. Compared with the single view detection, the multi-view detection structure can achieve greater accuracy with less time cost, and can effectively complete the detection task of 3D shape defect targets in actual industrial production.