caffe毕业设计论文.docVIP

基于caffe的卷积神经网络模型的评估 摘 要 本课题主要依据caffe平台进行神经网络的训练，主要对于本组成员所设计的神经网络得模型进行的评估和测试，测试时使用的是MATLAB，依据caffe提供的MATLAB的接口，使用的是caffe神经网络的模型，模型分别为alexnet,caffenet,lenet,目的是测试模型在新图片的准确率，同时对模型的网络文件的参数进行图像可视化，本课题通过研究三种模型，设计了算法来区分它们的识别率，测试识别率，模型参数可视化，测试准确率时是通过批处理测试，对比训练图片的标签，和测试图片得到的标签来进行判断：在对于模型进行参数可视化时，对第一层的卷积三个通道进行参数可视化，本课题在测试模型对识别图片的准确率上，并未采用单张图片进行测试，而是使用了批处理测试，首先对图片进行预处理，将图片转化为lmdb格式，最后得到均值文件，再进行全部的测试。在可视化时，将原来的3*num_output的卷积层参数视图，转化为正方形的图像，最后在比对图像的清晰度。本课题最后得到图像分别为alexnet，caffenet,lenet三种神经网络的参数可视化图像，得到的caffenet的图像应该为清晰的灰度图像，准确率方面也应该比其他的两种模型测试得到的结果要高。 关键词：caffe，神经网络，参数可视化，卷积层 ENGLISH SUBJECT ABSTRACT Neural network based on the Caffe platform, training, evaluation and testing for the design of neural network for the members of this group, Caffe provides MATLAB interface based on the Caffe neural network model, models were alexnet, caffenet, lenet, test the accuracy of the model in the new picture, at the same time parameters on the model of network file image visualization, this paper through the Study on three kinds of model, the algorithm is designed to distinguish them from the recognition rate, the recognition rate of the test, mainly through batch test, comparison of training images and test images from the label, label to judge, after the parameters in the test model of visualization results, also designed the algorithm, the main use of the Gabor function, it is to extract the features of the images, this test by using the Gabor function for each filter feature extraction and visualization of convolution of the first layer parameters of three channels, the subject in the test model on the accuracy of image recognition, did not use a single image for testing, but the use of the first batch of test. Preprocessing the image to a picture into a LMDB format, and finally get the average file, then all the test. In the visualization, the original 3*num_output convolution layer parameter view is transformed into a square image, and finally the definition of