Research And Implementation Of Multi-Exposure Image Fusion Algorithm Based On Deep Learning

With the continuous development of digital imaging technology,digital images have become an indispensable part of people’s daily life because of their convenience of storage.However,compared with the visual perception images acquired in real scenes,the digital images currently available do not meet the urgent need for high definition images.Multi-Exposure Fusion(MEF)can quickly and efficiently bridge the dynamic range gap between real scenes and image acquisition devices to obtain more realistic and vivid simulation of real scenes,which is beneficial to human eye recognition and computer processing.This paper researches and implements the multi-exposure image fusion algorithm based on deep learning technology,with the intention of improving the quality of multi-exposure image fusion,and the main research contents of this paper include the following aspects.(1)An efficient extreme exposure image fusion method is proposed.For the traditional convolutional neural network-based multi-exposure image fusion algorithm in a set of multi-exposure images for training,verification or testing,choose all the images of different exposure levels as the input of the model,with the deepening of the network model,which poses a certain challenge to the computational and storage capacity of the server,the waiting time for the output results is also longer.It also requires that one should not move significantly while performing the image capture,otherwise one will get a very bad image.To address this problem,the fast extreme exposure image fusion algorithm proposed in this paper selects only two extreme exposure images as input,adopts a lightweight network model to reduce the network running time,and introduces a new SimAM attention mechanism to efficiently extract image features and ensure the visual quality of the fused images.(2)For the fast extreme exposure image fusion method in the network training data and reduce and image size too large problem,by the random segmentation of the image,the segmented image for random region of Gaussian blurring processing,but also inversion,rotation and other operations to expand the data set,so as to achieve a sufficient number of multi-exposure images used for the training of the proposed network.(3)In order to improve and enhance the network model’s ability to acquire and reconstruct image feature information at different scales,a multi-exposure image fusion algorithm based on improved U-net multi-scale attention is proposed.The new improved multiscale module is introduced to cooperate with the SimAM space-channel dual attention mechanism in the fast fusion method to process different operations of the predecessor feature information of the downsampling operation and the hopping link operation in the U-net network.The experimental results show that the proposed method improves the quality of the fused images and outperforms other multi-exposure image fusion methods while equal exposure image input and no increase in network running time.