The Optimization Of Image Representation With Feedback Based On Deep Learning

With the rapid development of large data industry, people will have a lot of data every day via the Internet, the information is in a growth spurt in the development stage. So in the context of such a huge amount of data, how to quickly get the information we obtain useful information becomes particularly important. For the traditional text information, we can use the keyword lookup mechanism, but for image information, the traditional methods have not apply keywords. With the development in recent years, the deep learning, especially convolution neural network is widely used in the field of computer vision, image information retrieval problems began to be solved, and in the fast developing.This thesis presents a neural network by the deep of convolution method, the map image information to a method of feature space. The use of images in the feature space vector information, multi-vector classification of image problems. The innovation of this thesis is based on the deep of convolution neural network above, use triples loss function, in order to explore data distribution on the image expression, to seek within the warranty fuzzy class distinction clear class difference between the best goal The sample combinations. Compare the differences between the different loss functions. CAFFE frame as the current mainstream deep learning framework, which in industry and academia has been widely used. The advantage is convenience model definition, data engines common used, papers using multiple advantages. In this thesis, the use of CAFFE framework to build a model, and completed mentioned here in innovation content of the code written in the frame basis.In this thesis, a stochastic gradient descent procedure, in a grouping being reorganized data in the feature space, seeking to maximize the minimum distance of the sample target sample, reasonable to avoid the inherent similarity between samples. Among the tests, the use of the MNIST handwritten data sets, which is characterized by easy classification, and thus take advantage of the network of the two-dimensional feature extraction, we conducted a visual display of data distribution. Finally, it is also utilized Ali Taobao commodity dataset for image classification, in order to achieve good results. For the evaluation of the data set, we use the ROC curve and AUC as the evaluation index system, the different models of the quantitative evaluation. And for MNIST datasets show the spatial distribution of intuitive features.