Convolutional Dictionary Learning Using Global Matching Tracking (CDL-GMT): Application to Visible-Infrared Image Fusion

Abstract

The traditional convolutional dictionary-learning algorithm not only realizes the global sparse representation by imposing constraints on the image □0 or □1 norm, but also allows all possible movements of the local dictionary. However, selected atoms may be concentrated in certain areas of the image, while other atoms may be very sparse. Therefore, when using a traditionally-learned convolution dictionary, global sparseness alone is not sufficient to represent the entire image structure, and the resulting fusion image suffers from partial detail damage. For the above-mentioned convolutional sparse coding problem, this paper presents a greedy strategy based on the constraint 1_("0," ∞) problem to obtain a convolution dictionary(CDL-GMT), and applies the learned convolutional sparse dictionary to infrared and visible-light image fusion. This method attempts to impose constraints on sparsity locally, while considering the global structure. Experimental results prove the method to be superior to others in subjective and objective evaluation.