Design And Implementation Of Image Retrieval System Based On Deep Neural Network

Image retrieval helps humans gain the objective visual information.It has a broad range of applications such as portable devices using the technique of content based image retrieval.In order to improve the quality and speed of content based image retrieval task,an image retrieval system using deep neural network is proposed.The main work and contributions of this thesis are introduced as follows:A formal description of the image retrieval task is described and a convolutional neural network(CNN)model based on image hash retrieval algorithm is developed.And this thesis also applies indexing experimental data and evaluating methods.Location sensitive hash,spectral hash and deep autoencoder hash are implemented and optimized for collected dataset.This thesis compares these three algorithms on clothing image dataset.From the results,it indicates that the deep autoencoder hash outperforms the other two.The system this thesis proposed is an image retrieval one for clothing.It uses both labeled and unlabeled data to train unsupervised deep autoencoder model on image features extracted from supervised CNN model.This system can be divided into four layers,presentation layer,logic layer,application layer and data layer,which each part applies stable open source techniques.This thesis evaluates it for the effectiveness.From the results,it shows that this system exceeds the current systems such as Taobao and Baidu Clothing Retrival.