MSCNNLP: A Multi-scale Convolutional Neural Network and Laplace Pyramid Method for Multi-exposure Image Fusion

Abstract

ABSTRACT Image fusion is an image processing technology that makes full use of the complementarity and redundancy of images and fuses them through specific fusion rules to obtain images with better visual effects. The fusion image can not only highlight the object information, but also retain the texture details of the surrounding environment. Aiming at the problems caused by blurring edges, details loss, reduction of contrast and clarity of image in traditional multi-exposure image fusion methods, we propose a multi-scale convolutional neural network (CNN) and Laplace pyramid method for multi-exposure image fusion in this paper. The source image is input into the region Laplace pyramid for decomposition. In order to preserve more detailed information and make parameters adaptive, the convolutional neural network is modified, which is used to generate the optimal weight graph to guide the fusion process. Finally, the fusion image is generated by inverse process. Experimental results show that compared with other fusion algorithms, this proposed algorithm improves the image contrast, retains the edge and detail information in the source image. Furthermore, the fusion results have better objective evaluation value.