Research On Clothing Image Retrieval Algorithm Based On Deep Learning

In recent years,e-commerce has been booming in the fashion field,and the amount of fashion image data on the Internet is increasing day by day.In order to process massive amounts of clothing image data,effective content-based clothing image retrieval becomes extremely important.Hashing methods,which represent images in binary code and use Hamming distance to judge similarity,are widely accepted because of their advantages in storage and search speed.Encouraged by the latest developments in Convolutional Neural Networks(CNN),this paper proposes an effective deep learning framework to quickly generate binary hash codes for clothing image retrieval.This paper proposes a new supervised deep hashing method for learning compact hash codes to perform content-based image retrieval.This target hash code is used to describe the relationship between different image contents.Then,the target hash code is fed to the deep network for training.For large-scale clothing image database,a clothing retrieval model based on deep network is proposed.After training,our deep network can generate hash codes with larger Hamming distance for images with different contents.Experiments conducted on standard image retrieval benchmarks show that our method is superior to other latest methods,including unsupervised,supervised and deep hashing methods.This article mainly completed the following work:The first is to build a clothing data set.There are a total of 104,000 clothing images in the clothing data set.The data set contains clothing category information,clothing boundary box position information and clothing mask information.The experimental results show that position key point information can overcome clothing lighting,deformation,and occlusion.The influence of mask information can improve retrieval accuracy.Secondly,the deep neural network is used to extract the depth features of the clothing image,and the traditional algorithms for extracting clothing features are compared and analyzed,such as global features(HSV features)or local features(SIFT features).Use classification accuracy to verify the advantages and disadvantages of feature extraction algorithms.In addition,in order to eliminate the interference of background factors,this paper uses the regional convolutional neural network to locate the clothing area,and compares with the traditional clothing positioning algorithm to improve the final retrieval accuracy.Finally,this paper compares different hash algorithms and proposes a deep hash network that combines a deep network and a hash algorithm.The network uses the calculated target hash code for training.Compared with deriving the hash code from the middle layer of the deep network,the method in this paper generates the hash code from the output layer.Compared with the training process of the pair and triple method,which requires two images or three images as a training sample,the training time is greatly shortened.After training,our deep network can generate hash codes with larger Hamming distance for images with different contents.Experiments conducted on the clothing image database collected in this paper show that our method is superior to methods including unsupervised,supervised and deep hashing.