Multi-Source Remote Sensing Image Registration Based on Scale-Invariant Feature Transform and Optimization of Regional Mutual Information
Zhao Liao-Ying,Lu Bu-Yun,Li Xiao-Run,Chen Shu-Han
DOI: https://doi.org/10.7498/aps.64.124204
IF: 0.906
2015-01-01
Acta Physica Sinica
Abstract:In order to further improve the precision of remote sensing image registration, we propose a new registration scheme by combining the scale-invariant feature transform (SIFT) and the optimization of regional mutual information in this paper. Firstly, taking advantage of the randomness and ergodicity of chaotic sequence, we present a new chaos quantum-behaved particle swarm optimization (CQPSO) algorithm to solve the premature convergence problem of the quantum particle swarm optimization (QPSO) algorithm. By taking full account of the quantity differences among the values of different dimensions for the particle location information, small disturbances are generated as the Hadamard product of chaotic sequence and the particle location information. Before being added to the particle location information, the small disturbances are adjusted by an evolutionary parameter to ensure that each new particle location information is within the scope of reasonable evolution. The image registration scheme consists of two processes, namely the pre-registration process and fine coregistration process. The pre-registration process is implemented by the SIFT approach with a reliable outlier removal procedure. By the repetitive fine-tuning of several selected matched feature point coordinates, a series of registration parameters is estimated by a least square method and used to construct initial particle swarms. Next, the fine coregistration process is implemented to obtain the optimal match parameters by maximizing regional mutual information based on CQPSO. The proposed CQPSO algorithm is tested on several benchmark functions and compared with QPSO as well as standard PSO experimentally. Furthermore, comparative experiments are carried out on the registration of remote sensing images with different ground resolutions and the registration of remote sensing images at different phases by using four algorithms: the SIFT algorithm, SIFT combined with PSO algorithm, SIFT combined with QPSO algorithm, and SIFT combined with CQPSO algorithm. The regional mutual information, root mean square error, and the joint histogram are used to evaluate the performance of the algorithms. The experimental results verify the superiority of CQPSO and the effectiveness of the proposed registration scheme.