Deep Learning Adaptive Sturdy Guided Filtering for Artifacts Removal in Infrared and Visual Image Fusion

Abstract

In computer visualization applications, Infrared (IR) and Visual (VIS) image fusion has been developed rapidly as an interesting research field with the intention of combining the relevant information and producing a new image with better smoothing, sharpening and edge preservation. For this purpose, a Two-Scale Decomposition using Phase congruency and Sum modified Laplacian with Adaptive Sturdy Guided Filtering (TSD-PS-ASGF) technique has been proposed to construct the weighted maps by considering the bright and spatial variations between the respective pixels. As well, the fused image quality was adequately sharpened and smoothed by using ASGF method. However, this technique does not prevent the artificial structures and blur information around the salient features. Also, it has high computational complexity for constructing the saliency maps of base and detail layers. As a result, in this paper, a TSD-Deep Neural Network (DNN)-ASGF technique is proposed that uses DNN instead of PC and SML methods for constructing the saliency maps. At first, the TSD method is applied to decompose the source image into base and detail layers. For base layer, a weighted-averaging strategy is used for fusing the base information. For detail layers, DNN is used that extracts the multi-layer features. By using these features, -norm and weighted-average strategy are used for creating many candidates of the fused detail information. Once these candidates are obtained, the max-choice approach is performed to get the final fused detail information. Then, fused base and detail information are combined to reconstruct the final fused image with the minimum computational complexity. This TSD-DNN-ASGF not only preserves, smoothens and sharpens the information of both IR and VIS source images, but also prevents the artifacts, artificial textures and blurred details efficiently. At last, the experimental results illustrate that the proposed TSD-DNN-ASGF based image fusion technique achieves more efficiency than the state-of-the-art techniques like TSD-PS-ASGF in terms of objective evaluation and visual quality.