Through the idea of contrastive learning, the neural network structure of two Dncnn channels is designed to remove the mixed structural noise of infrared images. Pre-training is required in the training phase, and contrastive training is fine-tuning, and the entire model is more dependent on pre-training results.