Abstract: A semiconductor defect recognition method based on Bi-2DPCA and improved convolution neural network is proposed in order to solve the surface defect problem of semiconductor devices in packaging process. In view of overcoming the problem of low recognition accuracy caused by uneven samples, the training image is transformed to construct virtual samples, and then the image is compressed with Bi-2DPCA to extract the main features of the image. The improved AlexNet network is used for defect recognition and classification. In order to solve the problem of low adaptability of the model caused by the diversity and unpredictability of the diode plastic sealing surface, a normal random sampling layer is proposed, which is added to the convolution layer of the AlexNet network for lower sampling. At the same time, DropConnect is introduced in the full connection layer to improve the generalization performance of the network.Experiments show that the proposed algorithm has a higher recognition rate than the related algorithm, and has been verified on the SMA surface image set. At the same time, the algorithm has a good generalization performance.