Exact Clothing Retrieval System Based On Deep Neural Network

Clothing image retrieval is a branch on image retrieval which is extremely close to common applications and is becoming more and more important with the popularizing of e-commerce.Specifically,the user takes or uploads a clothing photo as query and the clothing image retrieval system returns a ranked list of images retrieved from a large corpus,which contains the same or similar clothing.The main difficulty of system is that there are large visual differences between the query clothing image and clothing product photos in online shops including illumination,pose variance,scale changes and occlusions.The traditional image retrieval methods based on manual-crafted features do not work well on clothing image retrieval task which is sufficiently challenging.To address this problem,we take advantage of deep convolution neural networks which have obtained great success in domain of computer vision in recent years and design a specific algorithm to solve the clothing recognition problem.The main contribution of our works are:1.As deep learning can learn new effective feature in training data for new application,we collect a large-scale clothing image data set which contains nearly 800,000 labeled images to train deep neuron network.Experiments demonstrate that the feature extracted by deep neural network which is fine-tuned on training set can represents clothing images affected by various factors of circumstances more exactly and robustly.2.On the basis of the aforementioned deep neural network which learns the generic feature of clothing image,we propose a clothing image retrieval method based on clothing image's deep learning feature and colorfeature.Experiments demonstrate that our method outperforms traditional image retrieval methods based on manual-crafted feature.3.We applied the aforementioned clothing image retrieval method into practice.An online recommendation system of clothes in videos was implemented.During the video is playing,the system would dynamically recommend the clothing product on screen in real time to guide users who are interested in to learn and even purchase those products.This dissertation is focused on practical clothing retrieval application.A specific and feasible retrieval method based on deep convolution neural network is proposed,which has practical value and guiding significance.