Gated Fusion of Infrared and Visible Light Images Based on CNN

Abstract

As a new research direction, fusion image technology has attracted more and more attention in many fields. Among them, infrared image and visible image, the two kinds of multimodal data have strong complementarity, the fusion image of the two modes contains not only the radiation information of infrared image, but also the texture detail information of visible image. In this paper, a convolutional neural network-based encoding-fusing-decoding network model structure is used. In the encoding stage, Dense Block, which has the advantage of feature extraction, was adapted to extract the image features. In the fusion stage, four fusion methods were compared and analyzed, and the gated fusion was selected as the main method of fusion layer. In the decoding stage, RDB (Residual Dense Blocks) was used to restore the fused features to the fused image. The fusion image based on this method is sensitive to temperature characteristics and has a better performance in image quality. The fused image has a high contrast, a relatively smooth fusion effect, and the overall visual effect is more natural.