Imaging systems with different imaging sensors are widely used in the surveillance, military, and medical fields. Infrared imaging sensors are widely used because they are less affected by the environment and can fully obtain the radiation information of objects, but they also have the characteristics of being insensitive to the brightness changes in the visual field and losing color information. The visible light imaging sensor can obtain rich texture information and color information but will lose scene information under bad weather conditions. Pseudo-color of infrared image and visible image can synthesize new image with complementary information of source image. This paper proposed a pseudo-color deep learning method for infrared and visible images based on a dual-path propagation codec structure. Firstly, the residual channel attention module is introduced to extract features at different scales, which can retain more meaningful information and enhance important information. Secondly, an improved fusion strategy based on visual saliency is used to pseudo-color the feature map. Finally, the pseudo-color results are recovered by reconstructing the network. Compared with other advanced methods, our experimental results achieve the satisfactory visual effect and objective evaluation performance.
|