Research On Convolution Neural Network For Clothing Classification

With the development of information society,security and fashion has become indispensable.Describing a person's clothing on many occasions is a very important task.For example,a criminal investigation usually involves looking for a suspect based on a description provided by an eyewitness or searching for images captured by a surveillance camera.How to effectively in the security requirements of higher or more sensitive areas of reliable automatic identification or confirmation of identity is very important.This involves the detection of people,clothing classification,object detection and image classification is a very popular research branch of computer vision,and has a very broad application prospects,image classification of this subject has been studied in many ways has been made great progress,as a specific application of image classification,clothing classification research is still relatively small,the main work of this article is also around this direction to explore research.The main work of this paper is as follows:First of all,by understanding the principle of Sift features and the introduction of Dense SIFT features defects,Dense SIFT features introduced on the basis of the head image detection method,the use of SVM linear classifier head model features training model.In this paper,the Dense SIFT feature is used for the first time in the detection of human head and compared with the existing methods.Secondly,we can locate the layout space of the character clothes by the head position of the characters.Through the understanding of the image segmentation algorithm and the comparison of different edge detection operators,we finally select the Canny edge detection operator to realize the accurate positioning of clothes position in the image,And complete the division of clothes.Finally,using the convolutional neural network to classify the clothes which have been segmented,AlexNet is used to initialize the network parameters and fine tune the network parameters with the dataset,which reduces the training time of the network to a certain extent The overfitting of the network is avoided.Compared with traditional feature extraction algorithms,CNN has made great breakthroughs in experimental precision,and traditional features have irreplaceable advantages.We further integrate figure feature and CNN.Taking into account the clothes color,texture,shape features,we can affirm that the further integration of a variety of traditional features is worth further study.Future work will include building a personal costume data set.If we can construct a more complex data sets,we can validate the effectiveness of the proposed segmentation algorithm.In addition,it is necessary to propose a feature fusion algorithm which is more suitable for the character of clothing classification.