Research On Classification And Retrieval Of Clothing Images Based On Deep Learning

Nowadays,with the increasing challenge for high-precision apparel image classification and retrieval technology based on convolutional neural networks(CNN)have received largescale attention.It has developed based on multi-layer neural networks and it's mainly used by deep learning methods.This research has devoted to exploring the process of feature extraction of apparel images by the CNN,and then explore the classification and retrieval technology of different apparel images based on different CNN.The clothing image classification research is divided into two parts: The first part constructs five different neural networks based on individual task learning which are conventional neural networks,CNN carry-out foundation modules,including inception modules and residuals.CNN of the difference module and two types of migration neural networks,the second part constructs CNN based on hidden layer sharing and the partial derivative is hidden layer sharing based on multi-task learning.In the apparel image retrieval research,the pre-training model Inception V3 is used as the feature extractor and the three different levels of feature mapping,medium and deep in Inception V3 output,the fusion of shallow and deep feature are used for mapping.The feature data of the image set-ups a costume image feature database,then uses different equivalently unit algorithms to measure the similarity between the image to be retrieved and the database typical image.Based on the Tensor Flow GPU platform and different apparel image data set this research completed the above convolution neural network design,training,and testing of verifies.The five convolution neural network models based on single task learning intend to design in this research can achieve better abstraction the results show that the neural network model designed by using transfer learning.It has higher classification accuracy but it has more parameters the classification effect of the two convolution neural network models based on multi-task learning is not satisfactory.There is a large space for improvement,based on the feature data output by the first theory module of Inception V3 and the retrieval algorithm of the Bray–Curtis dissimilarity retrieval effect be good.