The development of image recognition technology has made lots of opportunities to the manufacturing industry, speed up the intelligentization of manufacturing systems such as product quality assurance, automated assembly, and industrial robot control. In IC industry, manual inspection of industrial products for flaws is costly and inaccurate. Therefore, technicians have researched and developed computer vision technology and applied it to defect detection. However, most of the existing CNN models are only aimed at a certain dataset, and the effect is not good for the mixed data set. In the paper, we use machine vision to identify different classes of defects and imperfections. Specifically, an improved model based on UNet and SegNet is proposed for flaw detection of cables and transistors. This paper starts with the traditional SegNet model, integrates skip-connection and Atrous Spatial Pyramid Pooling (ASPP) to improve the performance of the model, and integrates a 13-layer convolutional neural network (ECON) in the experiment for classification to improve the model’s performance. Accuracy. A dataset of electronic component images from industrial production is used to compare the improved model, SegNet and UNet, and consider the performance of the combined classification model. The results show that the combined ECON and improved models have higher accuracy in the confounding of the two datasets compared to other networks.