The Research And Application On Clothing Style Image Recognition Based On Multi-Feature Fusion

With the advent of the Internet, online shopping has become a popular trend, more and more people choose this new shopping patterns. At present, the only way of commodity search in most e-commerce website is the text search which is much subjective. In view of the clothing retrieval, consumers are more likely to search clothing using clothing image. With the development of machine learning, the clothing image retrieval based on image recognition has become a hot spot in recent years. Using the appropriate method of feature extraction and classification, the clothing style which consumers want has been identified accurately. It will be more objective to express the consumers'needs.This paper will study how to make use of image classification identification technology in order to make clear classification in the clothing style. For the clothing image, single feature cannot distinguish between different styles of clothing. The clothing has many different features, such as color, texture, shape and other low-level information. This paper will solve the problem of multi-feature fusion. Considering the particularity of clothing image database, we will research two kinds of classification method which are the SVM and the convolutional neural network, and apply in the clothing identification system.To solve above problems, researched include the following:(1) In order to solve the problem of multi-feature in clothing recognition, the clothing feature can be extracted from two aspects:the structure feature and statistical feature. Multi-step SVM (Support Vector Machine) classification model will be established based on the multi-feature fusion in feature-level including fuzzy and accurate classification model. The parallel fusion methods based on contour and texture feature and grid and density of local points feature are used as the input vector of two-layers SVM classification.(2) For the large quantities of clothing image database, CNN (Convolution Neural Network) classification model is suitable. We will research traditional CNN and GoogleNet network architecture and design a 17th floor GoodNet network architecture. The experiment contrasted different network (CNN) can prove that the GoodNet network in JDNet database can obtain higher recognition accuracy and efficiency.(3) We will research the weight solution strategy in decision-level fusion and add the decision-level fusion in the layer of softmax in GoodNet. The strategy of solving weights of particle swarm optimization was proposed. This strategy has less parameter, fast convergence speed and easy implement.