Infrared Image Recognition based on Hierarchical Attention Perceptron Model

Infrared image recognition is one of the key technologies in intelligent substations. Compared with tradition image processing methods, deep-learning algorithms are preferred due to their stronger modelling power. However, few training samples impair the model performance, and brings Transformer network into consideration. Transformer network utilizes context information such as object relations in an effective way, but the computation complexity is not controlled for dense predictions on high-resolution pictures. To solve this problem, a Hierarchical Attention Perceptron (HAP) module which conducts window-based self-attention manipulations is introduced. This local self-attention strategy is further enhanced by shifted window partition configuration which connects neighboring windows to provide more information. By replacing the original backbone network and encoder with HAP module, the Transformer network has higher recognition accuracy and lower computation complexity. Experiments based on field data help find the best architecture, and prove the effectiveness of the proposed method with the comparison to mainstream models.