Research On Target Recognition Based On Feature-level Image Fusion

Object recognition based on multi-sensor feature-level fusion is the important researching subject in computer vision. For improving the recognition accuracy of multi-sensor imaging systems, solving the difficult problems of feature-level fusion in multi-sensor systems, we research the feature-level image fusion for Object recognition based on multi-sensor.We study the algorithms of feature-level fusion based on the viewpoint of dependence and independence multi-variant data analysis, respectively. First of all, it is difficult to using traditional principal component analysis in multi-sensor systems, two direction complex valued principal component analysis (2DCPCA) feature-level fusion algorithms is presented, the method extracts PCA features from row and column directions. The experimental results show that 2DCPCA could get higher recognition accuracy than single sensor with PCA without the situation of exquisite illumination and pose changes. Secondly, we introduced the kernel canonical correlation analysis into feature-level fusion method, because traditional linear canonical correlation analysis (LCCA) does not effectively describe the non-Gaussian distribution data. Our experimental results show that the improved method has excellent performance to deal with illumination and pose changes and increase the recognition accuracy 5-10 percent than traditional methods and is better than PCA at the recognition accuracy. Lastly, we discuss the method of extract feature from the viewpoint of data independence and propose the feature-level fusion method based on complex independent component analysis (Complex ICA). The experimental results show that our method is optimal with little samples. It is also demonstrated that our method could improve the most discriminating one without the situation of illumination and pose changes.Furthermore, we introduce multi-feature co-variance matrix into target recognition。However, the distance measure of co-variance matrix in non-Euclidean space usually lead to singular solution and bad recognition accuracy, we put forward the normalized fisher linear discrimination. The experiment results show that our method can improve the recognition accuracy 20% for some targets and has good resistance to target image aberrance. We also study feature fusion based on natural computation. First, we discuss the convergence of particle swarm optimization and artificial immune system. Our mathematic demonstration shows that particle swarm optimization is not the absolutely convergence method, but it is the excellent optimization. Second, we give the principle of binary coding method for feature level. The different discrimination function's discriminalities are compared with each other. The experiment results show that our method halve the feature dimensions and improve the recognition accuracy, stability and robustness.