Research On Clothing Image Classification And Retrieval Based On Deep Learning

In recent years,the classification and retrieval technology for clothing image based on convolution neural network(CNN)have attracted plenty of attention with the increasing demand for high-precision clothing image classification and retrieval.CNN is one of the deep learning methods which is developed on the basis of multi-layer neural networks and is mainly used for image classification and recognition.This paper first explores the process of layer-by-layer abstraction of clothing image features by CNN.Then the clothing image classification and retrieval technology based on different CNNs were studied.Image classification research work is divided into two parts.In the first part,five different CNNs are designed based on single-task learning method.These networks are the conventional CNN,the CNN containing inception module,the CNN containing inception module and residual block,two transfer learned CNNs.In the second part,the hidden layers-shared CNN and the partially hidden layers-shared CNN are designed based on multi-task learning method.In the study of clothing image retrieval,the pre-trained model Inception-v3 is used as the feature extractor.The different levels of feature mappings of clothing image given by Inception-v3,which include low,medium and high level feature mapping and the fusion of low and high level feature mapping,are used as image feature data to build the database,and then different similarity measurement methods are used to measure the similarity between the image to be retrieved and the sample image of the database.Based on the TensorFlow-GPU platform and different clothing image datasets,the design,training and testing of the CNNs mentioned above are completed.From the experimental results,the following conclusions are obtained.The features extracted by CNN have better abstraction than the traditional SURF algorithm.The five CNN models obtained by using single-task learning method can achieve better classification results,among which the transfer learned CNN models have higher classification accuracy,but the number of their parameters is larger.Two kinds of CNN models obtained by using multi-task learning method are not satisfactory,and they leave a lot to be desired.The retrieval algorithm which is based on the feature mapping output by Inception-v3's first Inception module and bray-curtis dissimilarity gives better result.