基于核稀疏分类与多尺度分块旋转扩展的鲁棒图像识别  

Kernel Sparse Representation Classification and Multi-Scale Block Rotation-Extension Based Robust Image Recognition Method

作　　者：匡金骏[1] 熊庆宇[1] 柴毅[1] 

机构地区：[1]重庆大学自动化学院,重庆400030

出　　处：《模式识别与人工智能》2013年第2期129-135,共7页Pattern Recognition and Artificial Intelligence

基　　金：国家自然科学基金项目(No.60975022);国家863计划项目(No.2008AA01Z148);黑龙江省杰出青年科学基金项目(No.JC200703)资助

摘　　要：针对在图像旋转或局部扭曲变形等复杂情况下的图像识别问题,提出一种基于核稀疏分类与多尺度分块旋转扩展的鲁棒图像识别算法.该算法首先对图像进行多尺度分块与旋转扩展,使得字典能近似测试图像局部的旋转扭曲与各种排列组合.为了增加字典类间稀疏度,改善系统效率,提出一种字典降维策略.通过核随机坐标下降方法高效求解核稀疏分类的凸优化问题,进而通过对比不同类对测试图像的重构误差完成图像识别.实验表明,与经典方法相比,文中方法具有更好的识别效果,对图像旋转或局部扭曲变形等复杂情况具有较好的鲁棒性.The random permutations and combinations of local images in image recognition tasks are complex problems. In this paper, an algorithm based on kernel sparse representation classification and multi-scale block rotation-extension （KSRC-MSBRE） is proposed grids are used to segment the training image, and the to solve these problems. Firstly, the multi-scale rotation-extended methods are applied to create a dictionary which adapts to the random permutations and the combinations of local images in test sets. To enhance the sparsity of the dictionary and improve the efficiency of the system, a new strategy is proposed to reduce the dimensions of the dictionary. Then, a kernel random coordinate descent method is proposed to solve the convex optimization problem in the KSRC. The experimental results show the proposed method has robust performance when dealing with the random permutations and the combinations of local images, and it outperforms other classical image recognition methods.

关 键 词：核稀疏分类 多尺度分块旋转扩展 图像识别 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]