Infrared and visible image fusion via dual encoder based on dense connection

Abstract

Aiming at the problems of information loss and edge blurring due to the loss of gradient features that tend to occur during the fusion of infrared and visible images, this study proposes a dual encoder image fusion method (DEFusion) based on dense connectivity. The proposed method processes infrared and visible images by different means, therefore guaranteeing the best possible preservation of the features of the original image. A new progressive fusion strategy is constructed to ensure that the network is better able to capture the detailed information present in visible images while minimizing the gradient loss of the infrared image. Furthermore, a novel loss function that includes gradient loss and content loss, which ensures that the fusion results consider both the detailed information and gradient of the source image, is proposed in this study to facilitate the fusion process. The experimental results with the state-of-art methods on TNO and RoadScene datasets verify that the proposed method exhibits superior performance in most indices. The fused image exhibits excellent subjective contrast and clarity, providing a strong visual perception. The results of the comparison experiment demonstrate that this method exhibits favorable characteristics in terms of generalization and robustness.