结合Parzen分类器和卷积神经网络的图像识别

Image recognition on parzen classifier and convolution neural network

针对传统卷积神经网络(Convolutional Neural Networks,CNN)中所用Softmax分类器方法无法有效兼顾识别准确率及识别效率的问题,提出采用Parzen分类器结合CNN的方法进行图像识别,特征提取方面,采用CNN提取图像的特征;识别算法方面,借鉴传统分类器的优势,提出采用Parzen分类器方法进行图像识别。使用MNIST、Pet数据集对该方法进行了测试。实验结果表明,Parzen分类器能有效地进行图像分类,与传统Softmax分类方法相比,在MNIST上的识别准确率提高了0. 73%,识别效率提高了20. 14%;在Pet Dataset上的识别准确率提高了1. 14%,识别效率提高了12. 94%。本算法具有较好的泛化性能。

In view of the problem that the Softmax classifier method used in the traditional convolution neural network(CNN)can not effectively take into account the accuracy and efficiency of recognition,it is proposed to use Parzen classifier and CNN.Image recognition,feature extraction,Using CNN to extract the characteristics of the image;With reference to the advantages of traditional classifier,Parzen classifier method is proposed for image recognition.The method was tested using MNIST,Pet datasets.The experimental results show that Parzen classifiers can effectively classify images.Compared with the traditional Softmax classification method,the recognition accuracy on MNIST has increased by 0.73%,and the identification efficiency has increased by 20.14%.The accuracy of the identification on the Pet Dataset increased by 1.14%and the identification efficiency increased by 12.94%.The algorithm has good generalization performance.