Research On Infrared And Visible Image Fusion Based On Learnable Grouped Convolution

The fusion of infrared and visible images has made significant progress as a valuable tool in image fusion.The infrared image and the visible image respectively obtain corresponding characteristic information according to a specific sensor and use the complementary information obtained between the sensors.Through a specific image fusion method,the feature information of the image is effectively fused,the infrared target and the visible outline of the image are highlighted,and finally,the fused image has more abundant and easily recognizable scene information in the fields of military,civilian,target detection and recognition It has crucial research significance and application value.This paper mainly focuses on the design method of generating confrontation networks,learnable grouping convolution,and some feature loss,further researching the infrared and visible image fusion algorithm,and verifying the proposed new model method combined with subjective and objective evaluation methods.First of all,the article introduces the imaging principles of infrared and visible sensors and the related theories of image fusion technology,sorts out three image fusion methods based on deep learning,and proposes a method that combines the generation of confrontation networks with learnable A fusion method combining grouping and convolution.The method is based on the generation of confrontation networks as the main fusion framework.In the model's design process,inactive perceptual features are used as the feature input of content loss to constrain the pixel intensity of the thermal radiation information of the infrared image and the texture of the visible image information.Besides,detail loss and target edge loss are added to the generator to retain the image's texture structure and target information effectively.Different fusion strategies are adopted for the model's detail layer for different source images,making the model a powerful expansion effect.Aiming at the most prominent real-time problem in this field,a new end-to-end network architecture based on a generative confrontation network is proposed to prevent the Batch Normalization(BN)layer from producing artifacts in the deep network's fusion image training process will affect the training stability.We take the dense residuals block as the fundamental network construction unit,make full use of each layer's characteristics,and reduce the number of parameters to a certain extent to achieve the effect of deep network supervision.A learnable grouping convolution replaces the ordinary convolution layer,which allows flexible grouping structure and produces better representation capability.Compared with ordinary convolution,grouping convolution can better trade-off between precision and speed.To analyze the effect on the model in different data sets,we experimented with two kinds of public datasets.The effectiveness of this model method is proved by qualitative analysis and quantitative analysis.The experimental results show that the model can generate high quality,rich texture of fusion image,real-time performance compared with other methods.