Near infrared facial expression recognition mainly relies on local features of lazy images. When the ex- tracted features are interfered with, the accuracy of facial expression recognition is low. Therefore, a new near-infra- red facial expression recognition method based on deep learning networks is designed. Relying on the local optimization and preservation method of the image to reconstruct the image structure information, the reduced dimensionality near- infrared facial image is obtained. The application point distribution model detects all key points on the face, extracts regions of interest for facial expression recognition, and constructs a facial expression classification and recognition model using a deep learning network architecture. By adjusting the parameters of the recognition model, the recogni- tion results of facial expressions are obtained. The experimental results show that the average Acc value of the proposed method’s recognition results reaches 0. 95, greatly improving the accuracy of near-infrared facial expression recogni- tion.