A New Image Fusion Method Based on PCANet

Abstract

In recent years, a simpler principal component analysis network (PCA Net) has been proposed. The network needs fewer parameters to be adjusted, which can solve the problem of long training time. However, because of the infrared and visible images obtained under different environments, the gray distribution of the images is different. If only one base layer fusion strategy is used, the generality of the algorithm may be reduced. To solve this problem, this paper investigates two different base layer fusion strategies. First, the visible dataset is used to train PCANet; second, based on the alternating guided filter to single-scale decomposition of the image to obtain the base layer and the detail layer,then Use the convolution kernel of the trained first layer network to extract the depth features of the source image, process it by kernel-norm to get the activity level map, and then construct the detail layer fusion weights; The base layer uses the two methods of maximum normalization of gray difference and weighted average to obtain the preliminary fusion map of the two base layers. By analyzing the methods in this chapter and other methods in simulation experiments. The QNCIE method in this chapter is 1.25% higher than the other methods. The QFMI is 150% higher than other methods. QP, QY and QCB are higher than other methods.