一种改进FSC算法在图像识别中的应用

Application of the Modified FSC Algorithm in Image Recognition

针对快速稀疏编码(FSC)算法用于图像特征识别时精度不高、训练速度慢的问题,在FSC算法的基础上考虑特征系数的最大稀疏化和特征基向量的最大表示性两个约束条件,提出一种改进的快速稀疏编码(MFSC)算法。在MFSC算法中,训练特征基向量时引入拉格朗日对偶函数以获得稳健的最佳特征;训练特征稀疏系数时采用特征符号搜索法以缩短训练时间。所用的测试图像选自Poly U掌纹数据库,对MFSC特征采用极端学习机(ELM)分类器进行图像特征识别。进一步与基于稀疏编码(SC)和FSC的掌纹图像识别结果进行对比,实验结果表明所提出的方法用于图像识别时的精度和快速性得到明显提高,具有一定的实用性和重要的理论研究意义。

Aiming at problems of the low recognition accuracy of image features and the slow convergent speed existing in fast sparse coding （FSC） algorithm, and considering two constraint conditions of the maximum sparsity of feature coefficients and maximum representation of feature bases on the basis of FSC algorithm, a modified FSC （MFSC） algorithm is proposed here. In this MFSC algorithm, to obtain robust optimal features, the Lagrange dual function is used to train feature basis vectors, at the same time, to reduce convergence time, the feature sign search method is used to train sparse feature coefficients. In test, all images are chosen from the PolyU palmprint database. For image features extracted by MFSC, the extreme learning machine （ELM） classifier is utilized to implement the recognition task. Furthermore, comparing the recognition results of MFSC with those of sparse coding （SC） and FSC, experimental results show that the recognition accuracy and rapidity of MFSC can be clearly improved.