Research On Image Feature Representation And Classification Based On Hierarchical Classification

With the development of Internet and computer, the image data shows the tendency of explosive growth. There is a large gap between the image physical manifestation and people who are familiar with the concept of information. It is proposed the great challenge in process capability of the algorithm and speed of classification. So the outstanding feature extraction algorithm and classification model is an important research in large-scale image processing.This paper mainly studies the mass in the image classification based on hierarchical classification problems. The increase number of images, the features of the extract the image type and the number would be huge, this to the project application and equipment produced a huge test. First of all, how can not only improve the classification accuracy but also reduce the characteristic number has become a key issue.Second, shallow network the extracted features is not rich, expression is not comprehensive, is the main factor affecting the classification results.it is particularly important to choose the better alternative network. Finally, according to the performance of different network model under the large-scale data choose to suit oneself classification learning model of sample set, not only can improve the accuracy still can reduce the manpower and financial resources. To solve these problems, this article has carried on the related research, the main work summarized as follows:(1) For the shallow network characteristics extracted enormous problems, research and shallow network feature extraction method is put forward, namely, the first to use SIFT algorithm, then using the Locality-constrained Linear Coding (LLC) to sparse the number of characteristics. Effectively decreases the number of the characteristics and increase speed.(2) Because of the traditional feature extraction and classification of shallow network time-consuming and modified algorithm is difficult, is put forward on the classification results using the DARTS optimization algorithm based on hierarchical classification, the information gain of each node is calculated so as to make the under arbitrary precision weighing out the most accurate results.(3) Shallow network feature extraction based on hierarchical classification is not rich, deep learning theory knowledge is studied, using convolution neural network(CNN) for image feature extraction and classification of learning. After constant convolution and the sampling, finally learn out the multilayer network stack and depth of network structure.(4) Using shallow and deep network to make test with two different database. The two types of database are extracted road bayonet video and on common ImageNet database. Comparing two kinds of test database to the performance of the two kinds of classification algorithm.(5) Design an APP on your phone, set up based on the hierarchical classification of image classification algorithm of convolution neural network system, verify the validity of this method.