基于CNN多层融合特征与Fisher准则的分类算法

Classification algorithm based on multi-layer fusion feature of CNN and Fisher criterion

针对目标物体的多分类问题,提出基于CNN多层融合特征与Fisher准则的分类算法。首先,应用卷积神经网络提取图像的各层特征;然后,通过穷尽搜索法确定各层特征融合的权值系数,得到多目标分类特征;最后,采用Fisher多分类准则,求出使模式具有最大可分性的最佳投影方向,实现目标分类。在ORL、Yale库上进行实验研究,分类准确率分别达到了97. 5%和97. 3%。结果表明,该方法能够解决模式多分类的问题,与传统方法相比有效地提高了识别能力,具有很好的鲁棒性。

Aiming at the multi-classification problem of target objects,a classification algorithm based on multilayer fusion feature of CNN and Fisher criterion is proposed. First,the convolutional neural network is applied to extract the features of each layer of the image;Then,the weight coefficients of each layer feature fusion are determined by the exhaustive search method,and the multi-target classification features are obtained. Finally,Fisher’s multi-classification criteria are used to find the best projection direction with the maximum separability of the mode and achieve the target classification. Experimental studies were carried out on the ORL and Yale face databases,and the classification accuracy rates reached 97. 5% and 94. 6%,respectively. The results show that this method can solve the problem of multiple pattern classification and improve the ability of face recognition compared with the traditional method,and has good robustness.