Research On Image Retrieval Algorithm Based On Deep Learning And Hash Code

Nowadays,with the expanding of Internet and multimedia industry,content-based image retrieval(CBIR),as an accurate and fast image retrieval method,is becoming more and more popular.In recent years,CBIR is facing the main problem of reducing the semantic difference between the low level feature and the higher one.Besides,a CBIR system needs to retrieve similar images from massive image databases quickly and effectively.These two problems correspond to the two main steps of CBIR,namely,image feature extraction and feature indexing.The research of this paper is centered on these two main steps.In this paper,the SS-VGG16 convolution neural network based on unsupervised learning,is proposed,which is applied in the phase of image feature extraction.There are three innovations in this module,which solve three problems: too much parameters of VGG16 model,the problem that VGG16 network can only input fixed size images,and the problem that supervised learning can not be trained based on unlabeled image database.This paper improves VGG16 based on classical convolution neural network To improve the classical convolution neural network VGG16,this paper proposes a lightweight S-VGG16 network structure based on the idea of Squeeze Net.The purpose improving is to reduce the network parameters because less parameters represent shorter network training time and faster image feature extraction.In order to solve the problem that traditional CNN can't extract the features of images of any size,we use space pyramid pooling(SPP)to replace the last maximum pooling layer in S-VGG16 network and derive framework named SS-VGG16,which enables images of any size to be accepted as input.Therefore,there is no need for image stretching,cutting and other pre-processing so we avoid image distortion.Considering the image retrieval of unlabeled data sets,this paper proposes an unsupervised transfer learning method to train SS-VGG16.This method utilizes the inner relation of similar image data and uses the mean of image features as training target.In order to prove the effectiveness of the proposed SS-VGG16 network structure,a supervised learning experiment based on MIT Places 365-Standard data set is conducted.Top-1 and Top-5 accuracy are used as evaluation criteria.The image retrieval performance of the proposed SS-VGG16 and the original VGG16 is compared,and the effectiveness of SS-VGG16 is proved.In order to prove the effectiveness of unsupervised learning method proposed in this paper,we use Paris 6K and UKBench data sets to conduct experiment.MAP and Top-4 are used as evaluation criteria respectively.The proposed method is compared with unsupervised learning methods in recent years,proving the superiority of unsupervised training method in image retrieval tasks.This paper presents an image retrieval algorithm based on depth feature,feature dimension pre-reduction and iterative quantization coding.Before the stage of feature indexing,this paper proposes a feature reduction method based on auto encoder neural network for data non-linear learning to achieve the purpose of feature dimension reduction.In the stage of feature indexing,an iterative quantization(ITQ)method is used to index the image features extracted from SS-VGG16 network model and encode them based on depth learning.ITQ method keeps approaching the minimum quantization error between features and hash codes with set bits to minimize quantization error.In this paper,Recall,Precision and Mean Accuracy(MAP)are used as the evaluation criteria of retrieval effect.The corresponding experiment is conducted on Caltech 256 image database.The experimental results demonstrate that the effect of the proposed image retrieval algorithm has the superiority over other mainstream algorithms.