Research On Infrared And Visible Image Fusion Algorithm Based On Deep Learning And Total Variation Model

As an important part of image processing,image fusion aims to combine image information obtained by different types of sensors,and mainly uses nonlinear processing to output a single robust and information-rich image.The fusion process can extract valuable information from the source image and fuse it into an image without introducing additional artifacts,so that people can better understand the scene information.As an important branch in the field of image fusion,infrared and visible image fusion has important applications in military,security monitoring,face recognition and other fields.Visible images have the characteristics of clear imaging,strong contrast,and rich information,but changes in external light,such as weather and other conditions,have a greater impact on their image quality;infrared images are suitable for low-illuminance or hidden target recognition,but their spatial resolution is low,insufficient texture information.Infrared and visible image fusion can fully combine the advantages of infrared and visible images to generate a fused image with rich texture details and clear targets.Existing studies have shown that the proposed fusion algorithm for infrared and visible images can make the target information of the fused image the same as the infrared image,and the background texture is the same as the visible image.However,there are still problems such as low contrast,blurred details,and loss of edge contour information.In view of the existing problems of the current fusion algorithm,the main research content and work of this paper are as follows:(1)Combining the relative total variation model that can separate the texture and structure information of the image as much as possible,and the characteristics of the least square filter to preserve the edges,a hybrid multi-scale image decomposition method is proposed to complete the image fusion task.The method of relative total variation is used to extract the detailed information of the image,and the base information extracted by the least square filter.The contrast saliency and gradient saliency mapping are used to construct fusion weights to process the detail layer and the base layer information separately.The proposed algorithm overcomes the problems of low contrast,insufficient edge preservation and unclear details in traditional fusion methods.(2)Combining image super-resolution technology and Retinex enhancement algorithm,an infrared and visible image fusion algorithm based on image enhancement technology is proposed.Specifically,the enhancement algorithm based on convolutional neural network is used to enhance the detail information and contour information of infrared images and visible light images,and the improved Retinex enhancement algorithm in this chapter is used to enhance the illumination brightness of low-light visible light images.This algorithm can improve the overall brightness of the fusion image and retain the clear outline and detail information in the original image.(3)Taking into account the advantages of the automatic decoder in the Densefuse network,but because the network adopts the weighted average fusion rule,this rule will weaken the characteristics of the fusion result and make the fusion image more blurred.Therefore,this chapter improves the network and designs an enhanced autoencoder.The autoencoder consists of two parts: an encoder and a decoder.First,the encoder extracts the basic features of the visible light image and the infrared image,then integrates the image features through the fusion module,and finally uses the decoder network to reconstruct the features into the fused image.This algorithm improves the shortcomings of the original Densefuse algorithm,and the resulting fusion image performs well in terms of brightness,contrast,and contour information.