Research On Infrared And Visible Image Fusion Algorithm

Infrared image and visible image are common information carriers in the field of image fusion.The integrated image has many advantages,such as significant target object,high contrast,rich background details and so on,which can significantly improve the image quality.Based on the study of existing image fusion algorithms,this paper proposes two infrared and visible image fusion algorithms.The main work is as follows:(1)Traditional multi-scale fusion methods use edge decomposition filter.Aiming at the problems of complex design,high time-consuming and insufficient detail information extraction,this paper proposes an infrared and visible image fusion method based on double-scale decomposition and potential low rank representation.Firstly,a simple and efficient averaging filter is used to achieve the purpose of multi-scale decomposition,and the infrared and visible light source images are decomposed into basic layer and detail layer;Secondly,the potential low rank representation is used for the basic layer to obtain the low rank layer and the saliency layer,so as to obtain more detailed information;Then,for the detail layer,visual saliency detection is carried out to obtain the saliency map,and finally the normalized weight map is obtained.The detail layer,low rank layer and saliency layer are fused by three different rules:weighted fusion,averaging and summation strategy;Finally,the fused image is obtained by multi-scale reconstruction.Experimental results show the superiority of this algorithm.(2)Aiming at the problems that the deep neural network framework is only applied to some parts of the fusion process,while the overall fusion process is still in the traditional framework,and the traditional fusion methods need to manually design the fusion rules,which is easy to lead to the loss of a large amount of information of the fused image,this paper proposes an infrared and visible image fusion method based on a dual-discriminator generation adversarial network.Firstly,the traditional generator is improved.The generator adopts double branch network to extract the detail information and semantic information of the source image respectively;In order to alleviate the gradient disappearance,repair the feature loss and reuse the previous features,dense connections are referenced on the detail branches.Secondly,the method adopts the double discriminator model to form a confrontation game between one generator and two discriminators.The first discriminator is to train the fusion to be close to the visible image,and the significance of existence about newD_i is to facilitate the discrimination between the generated image and the visible image.Therefore,in this network,the fused image is trained to be similar to two source images at the same time.Finally,in order to reduce the feature error between the fused image and the source image,this paper also introduces the perception loss function to further improve the fusion quality.Experimental results show the superiority of this algorithm.