The defect detection of wafers is an important part of semiconductor manufacturing. The wafer defect map formed from the defects can be used to trace back the problems in the production process and make improvements in the yield of wafer manufacturing. Therefore， for the pattern recognition of wafer defects， this paper uses an improved ResNet convolutional neural network for automatic pattern recognition of seven common wafer defects. On the basis of the original ResNet， the squeeze-and-excitation （SE） attention mechanism is embedded into the network， through which the feature extraction ability of the network can be improved， key features can be found， and useless features can be suppressed. In addition， the residual structure is improved， and the depth separable convolution is added to replace the traditional convolution to reduce the computational and parametric quantities of the network. In addition， the network structure is improved and the activation function is changed. Comprehensive experiments show that the precision of the improved ResNet in this paper reaches 98.5%， while the number of parameters is greatly reduced compared with the original model， and has well results compared with the common convolutional neural network. Comprehensively， the method in this paper can be very good for pattern recognition of common wafer defect types， and has certain application value.