Cross-Modal Image Matching Based on Coupled Convolutional Sparse Coding and Feature Space Learning

摘要

Coupled sparse representation (CSR) has achieved great success in cross-modal image matching in recent years. This model not only observes a common feature space for associating cross-domain image data for recognition purpose, but also improves the accuracy of cross-modal image matching. However, it applied the locally based manner in coupled sparse coding stage, so it is difficult to extract deeper image features, which is very important to improve the accuracy of cross-modal image matching. In order to address this problem, this paper proposes an iterative cross-modal image matching algorithm based on the coupled convolutional sparse coding and feature space learning (CCSCL). In contrast to the existing CSR-base cross-modal image matching methods, this model provides a global and flexible way to overcome some limitations of CSR. By making use of convolutional sparse coding and operating on the whole image, CCSCL can grasp the correlation between the pixels well and get more accurate modal feature map. In addiction, an cross-iterative training algorithm based on common feature space and correlation analysis of coupled convolution sparse coefficient is derived to efficiently solve the optimization problem. Experimental results show that the proposed model can effectively apply to cross-domain image matching and attain 98% matching accuracy.