Research On Apparel Retrieval Based On Semantic Segmentation

With the popularity of the Internet and mobile electronic devices,the traditional way of shopping is gradually replaced by online shopping,which has become a new trend.Among them,clothing,as a necessity,occupies an important proportion in online shopping.At present,most of the clothing shopping platform's clothing retrieval technology is based on the keyword retrieval method,but it can not fully describe the image content,manual annotation clothing attribute subjectivity is too large,affect the accuracy of clothing retrieval and other problems.The Content-Based Clothing Image Retrieval method can solve the above problems.The key of Content-Based Cothing Image Retrieval limits the feature extraction and feature indexing mechanism of clothing image.At present,there are many problems in the existing Content-Based Clothing Image Retrieval methods,such as inadequate use of clothing image data information,rough processing of target edge information,and too much calculation in the implementation of clothing data indexing.How to retrieve clothing efficiently and accurately in the massive clothing data has been a research topic.In order to the problems of inadequate use of clothing image information,rough processing of target edge information and excessive amount of calculation in clothing image retrieval,this thesis first designs a clothing image retrieval algorithm based on double Attention Mechanism to guide Deeplab V3+,which improves the extraction of adjacent information of clothing image,improves the processing of clothing edge information,and is suitable for clothing Index Mechanism On the premise of extracting the clothing feature vector before,the time of clothing retrieval is reduced.The main contents of this thesis are as follows:1)Aiming at the problems of poor performance of small target segmentation,insufficient feature information extraction and rough target boundary in clothing image,a DoubleAttention directed Deeplab V3+ model is proposed as clothing image segmentation network.Based on Deeplab V3+ network,this method uses feature pyramid instead of spatial pyramid,and combines with Attention Mechanism to fuse context information of adjacent scales of image,so as to provide more spatial attention information for clothing image segmentation.In order to verify the effect of the algorithm in this chapter,the model built in this chapter is trained and tested on two image datasets of Product Clothing Dataset and Daily Clothing Dataset,and several algorithms are used to compare the F of different types of clothing images.The experimental results show that: compared with the other two algorithms,the algorithm in this chapter has better segmentation effect.2)In view of the problem of how to reduce the time required for clothing retrieval under the premise that the clothing Index Mechanism is suitable for the current extracted global feature vector in the massive clothing image data set,a clothing image retrieval algorithm based on Double-Attention Mechanism and Deeplab V3+ Hash-Inverted Index is proposed.The network selects the commonly used inverted index suitable for massive data as the index But it is not suitable for the global feature vector extracted by Deeplab V3+ network.Therefore,this thesis proposes a Hash-Inverted Index method based on inverted table,which uses PKM algorithm to build a large visual dictionary on the basis of inverted table to make it suitable for global features.At the same time,in order to the quantization error between the features extracted by Deeplab V3+ and visual words,a binary feature is generated by using embedded code.However,it has the problem of repeated comparison of the same clothing image,so we propose to use hash code instead of embedded code to solve the problem of repeated calculation of the same image.In order to verify the effect of the algorithm in this chapter,experiments are carried out on the clothing image dataset of Deep Fashion Dataset,and different hash codes are used to compare MPA.The experimental results show that,compared with the index method of other length hash codes,the Hash-Inverted Index accuracy of hash code length 12 is higher.At the same time,compared with the original inverted index for retrieval time and MPA value,the retrieval time is relatively reduced,and MPA value has been improved to a certain extent.This method improves the precision and speed of clothing image data retrieval.