Reasearch And Application Of Clothing Attributes Based On Deep Convolutional Network

The research of Clothing attribute include recognizing the class of clothing style, detecting clothing parts and retrieving similar images on an image. Because of the limitation of hand-designed feature and the lack of public datasets, clothing attribute classification and retrieval problem is extremely challenging in practical application. We propose a model based on convolutional neural network. Our contributions can be summarized as:We propose a model based on deep convolutional neural network for clothing style classification and retrieval task. We first establish the B_Dat Clothing dataset containing more than one hundred thousand clothing images of different styles. We train deep convolutional neural network model on B_Dat clothing dataset and do validation on test set. We then extract feature from the seventh fully connected layer for all images and build hash indexing with iterative quantization to generate hash code. Experiments show that our method based on convolutional neural network and constructing hash indexing with convolutional feature can achieve good results with high accuracy of classification and fast retrieval.Clothing items detection model is mainly based on region convolutional neural network. We detect clothing items and human body with fast-rcnn and faster-rcnn. We merge the labels of street images into 12 categories (e.g. bag, belt, face). We refine the parameters of fully connected layer to achieve better results on clothing items detection and human body detection based on fast-rcnn and faster-rcnn. Comprehensive evaluations demonstrate that the accuracies of two models are similar, but the faster-rcnn method is faster and can detect objects in time. With the expansion of training samples, the accuracy will be higher. The method can locate the positions of clothing items on images with complex background, which will improve the accuracy of classification.The system applications were based on the attributes of the garment from clothing attribute classification retrieval of PC client, the clothing parts for real-time detection, web clothing retrieval. For these three aspects, this paper design and develop the different function software in the three different operating platforms. PC client system designed to run under the Windows operating system, which realized the users to upload a clothing picture and detected the attribute, color, and provided users with other functions including offline and online retrieval. The real-time detection for clothing parts may help users detect the parts of clothing and human. It would provide foundation for future research. The web clothing retrieval realized the interaction between users and system. The web fast retrieve similar clothing pictures by the users uploaded pictures in the local address or network address. The experimental results showed that the system application based on attributes of the clothing which achieved good performances in the three different operation platforms.For the weakness of traditional manual designed features in the classification and retrieval of clothing, the paper adopted the basic framework of deep convolution neural network (DCNN). The experimental results show that the convolution features in combination with a hash index, which could fast accurate classification and retrieval of similar images. The clothing parts target detection on the speed and precision has achieved good performances. The designed system software also has good performances in the reality experimental.