Research On Image Retrieval Based On Deep Learning

With the rapid development of the Internet and the wide application of multimedia technology,the amount of image data in the network is increasing.How to query the desired image accurately and quickly in a large number of image data is the key problem of image retrieval task.In recent years,as a representative nearest neighbor retrieval technology,hash algorithm has become the main method of current image retrieval tasks.Traditional hashing methods usually use manual image features to learn hash codes.The inadequacy of manual feature expression affects the retrieval effect.With the development of in-depth learning,retrieval methods based on in-depth hash network are becoming popular,but there are still some problems such as imbalance of training data,large quantification error and lack of semantic information.To solve the above problems,this paper proposes an end-to-end deep hash network architecture,which achieves high-quality hash codes through feature extraction and binary hash code generation.The main work of this paper is as follows:1.In order to extract image features with strong expressive ability,this paper uses VGG16 network and introduces migration learning.Firstly,VGG16 network is pre-trained on Image Net dataset,then it is migrated to image retrieval task,and fine-tuned by image retrieval dataset to adapt to image feature extraction in image retrieval task.At the same time,the last full connection layer of the network model is changed into hash learning layer,which ensures that the feature representation and hash code of the image can be learned while training the network.2.In order to obtain high-quality hash codes,this paper takes the similar label information of image pairs as the supervisory information in training,and preserves and learns the similarity between images,which makes the Hamming distance between similar image pairs and hash codes learnt through network smaller,and the hash code distance between dissimilar image pairs larger.At the same time,aiming at the imbalance of similar data in training,this paper introduces a weight coefficient to adjust the punishment of learning similar images and dissimilar images,so as to reduce the impact of imbalance.3.In order to reduce the quantization error caused by the generation of hash codes,this paper improves the activation function by adding a regular term to reduce the error causedby discrete optimization in the process of network learning.At the same time,aiming at the problem of missing semantic information,the global loss function is improved by combining the loss of soft Max classification and image contrast,which makes feature extraction and hash quantization feedback mutually.The end-to-end depth hash image retrieval algorithm is realized,which improves the retrieval accuracy.Finally,experiments are carried out on CIFAR-10 and NUS-WIDE datasets.In order to verify the effectiveness of the retrieval algorithm,a comparative experiment is designed to analyze the experimental results of the improved method.The results show that the proposed method effectively improves the retrieval performance and can query more suitable images.