Abstract: In order to improve the accuracy of the deep neural network in the diagnosis of pneumonia, an improved neural network model method based on recursive gated convolution is proposed, which realizes the high-order spatial interaction between features in a recursive perspective, and helps the network interpret the chest X-ray image data in a more global perspective. By using Taylor expansion to improve the polynomial coefficients, the loss function could accelerate the convergence speed and improve the accuracy and performance. At the same time, combined with a regularization function, the sample feature norm extracted by the neural network was used as the sample attribute for learning. According to the size of the feature norm, the sample distribution of sample recognition difficulty from simple to hard was realized, so as to avoid the neural network over-fitting low-quality noise samples during the training process, and improve the training efficiency and the rationality of the sample distribution. The experimental results show that the accuracy of the improved network model on the pneumonia dataset has improved by an average of about 1.5 percentage points compared with VGG11, EfficientNetV2_S, ResNet34 and other network models.