Infrared images can directly reflect the temperature of a circuit board and its changes. To address the challenge of accurately locating chips on infrared images of circuit boards, this paper proposes a segmentation method based on the GAN-SUNet model. The SUNet model is improved from the UNet model by introducing a spatial pyramid pooling (SPP) module, modifying the loss function, and reducing the number of convolution cores to enhance detection accuracy and network speed. First, a GAN is used to learn and train on collected infrared circuit board data to generate simulated infrared images and expand the dataset. Then, the generated dataset is used to train the SUNet model, and model parameters are adjusted to improve verification accuracy. Finally, the trained model is used to identify, detect, and segment chips on circuit boards, thereby achieving chip localization in infrared images. Experimental results show that for infrared image segmentation of circuit boards, the GAN-SUNet model achieves an average intersection and merging ratio of 93.77%, effectively reducing the burden of manual chip localization and providing strong support for subsequent chip temperature data processing.