Research On Fusion Of Infrared Image And Visible Image

Image fusion is mainly used to synthesize the complementary information of multisource images.Identify the same scene by generating fused images.The source images come from different types of image sensors,and the characteristics of the images obtained are different.The fused image will have all the advantages and characteristics of the source image,and will be more easily viewed by humans or recognized by machines.In recent years,with the development of digital photography,remote sensing image,video surveillance,medical image and other image applications,the field of image fusion in image processing has become increasingly hot.Fusion of infrared and visible images is an important research direction.Infrared images will distinguish the target from the background according to the radiation difference,and the image shows a clear outline of the object,which is not easily affected by weather light,etc.,and works well in both day and night.Compared with the global structure information of infrared image,visible image provides texture details with high resolution and clarity,and can see the rich content of image more clearly.However,it is easy to be affected by light and weather,and the structure of objects in the image is not as clear as infrared image.The fusion of these two images can combine the contour information of the object in the infrared image with the texture information of the object in the visible image.High quality fusion images have many practical applications.In this thesis,the fusion algorithm of infrared and visible image is studied.The introduction is divided into the following aspects,including infrared image and visible image characteristic analysis,image preprocessing,image fusion related methods,image fusion quality evaluation and image fusion field existing problems and hot spots.Aiming at some problems existing in traditional infrared and visible image fusion,including fuzzy image contour,unclear image texture details,and not obvious image features,two fusion methods are proposed,the main contents are as follows:1.Fusion method of infrared and visible image based on dense connected GANAiming at the simple convolution process of GAN(Generative Adversarial Networks)networks,an image fusion method based on dense jointed GAN networks is proposed.The dense connection is added to the fusion method to cascade the shallow information with more detailed features to the deep information,and the convolutional input of each layer is cascaded to all the convolutional output of the previous layer to generate more detailed information of the image.The extraction of visible light information is increased by adding the visible light source image again before convolution.The corresponding generator and discriminator are designed,and the experiment is regulated by reasonable loss function.A useful end-to-end fusion network was obtained through multiple training.The fusion image is generated by experiments.Compared with some classical fusion methods,the fusion image obtained by the proposed fusion method is clearer in human vision and contains more details and textures.In terms of subjective evaluation,the fusion image seen by human eyes will be clearer and richer than the infrared and visible light source images,and it is easier to see the key information of the image.In the objective evaluation,using the representative image evaluation criteria,the fusion image has better quality.2.Infrared and visible image fusion method based on three-branch autoencoderThe existing image fusion tasks are basically based on common autoencoder networks,ignoring the extraction of important features.To solve this problem,a unique autoencoder network is proposed,which is characterized by three branches,namely,structure branch,content branch and foundation branch.In the encoder,the important features of infrared image and visible image are extracted through three branches.The structure branch extracts the structure and contour information of the source image,the content branch extracts the main content information of the source image,and the basic branch extracts the basic information and features of the image in order to ensure the minimum distortion of the fused image.The addition strategy and channel strategy are used to fuse the extracted features to obtain the fusion features.The decoder gets the fused image by decoding the fused features.This method is applied to image fusion experiment,and good results are obtained.In the evaluation index also achieved a good result.