Research On Infrared And Visible Image Fusion Based On Generative Adversarial Network

Infrared and visible image fusion is one of the basic tasks in the field of computer vision,which plays a key role in the subsequent processing of images.It has been widely used in remote sensing image processing,target detection,security monitoring and military fields.However,the scene with poor illumination conditions will introduce more interference information into the fused image.Then,the simple structure of the generator network of the Fusion GAN method makes the extraction of salient features inadequate.Besides,the loss function only considers the information of visible gradient and infrared intensity,and other important information will be lost.In order to solve the above problems,a new image fusion method of infrared and visible based on generative adversarial networks is proposed.Firstly,to eliminate the influence of illumination environment on image fusion performance,a visible image preprocessing module is designed.Before the visible image is input into the generator,the dark region is enhanced to suppress the noise introduction.Secondly,dense convolution blocks are embedded after generator network Conv1 in order to make source image feature extraction more sufficient and alleviate the gradient disappearance phenomenon.Finally,in order to enrich the complementary information in the fused image and improve the quality of the fused image,structural similarity loss and pixel loss are added to the generator loss function.An optimized loss function is established to construct the end-to-end image fusion process.To verify the effectiveness of this method,comparative experiments are carried out on TNO dataset and KAIST dataset with traditional image fusion methods and deep learning image fusion methods.Experimental results show that the proposed method is superior to current image fusion methods in performance.In subjective evaluation,the visual effect of the proposed method is better than that of the original method.In terms of objective evaluation,the proposed method is better than the original method in the following indexes: information entropy,mutual information,difference correlation and average gradient are improved by 0.9183,1.8365,0.4612 and 0.2004 respectively.Better performance is achieved and the quality of fusion image is improved.There are 29 figures,14 tables and 54 references in this paper.