Research On Image Retrieval Based On Deep Learning

In recent years, with the progress of Internet and the multimedia technology, people are exposed to increasing network data. Image which is the main component of network data contains a large of information. A great challenge we faced is how to dig out quickly and accurately the image information that users needed from a large-scale image data. Organizing and managing effectively the image data is also difficult. In order to retrieve effectively desired images from the image resources, content-based image retrieval technology emerges. The technology has great significance and practical value, and it has become a hot topic at home and abroad. Although researchers have conducted a lot of research on content-based image retrieval technology and achieved a lot of research results, many technologies are not still mature and need to be studied further.The research present situation and problems of content-based image retrieval are first analyzed in this paper, and the basics of content-based retrieval and deep learning model are elaborated. After understanding and grasping the related knowledge of image retrieval, an image feature extraction algorithm based on deep learning combining supervisory signal is proposed in this paper. The algorithm can well express image feature which is not obvious. Faced with the problem that the classification and retrieval accuracy are not high enough in the current image retrieval, a stacked autoencoder network algorithm using hashing and two supervisory signals is proposed in this paper. The algorithm has strong classification and learning ability, and improves greatly the classification and retrieval accuracy. That is, the algorithm achieves better retrieval effect by reducing the difference between similar images and increasing the difference between different images. The algorithm is divided into two phases:(1)Firstly, the algorithm fine-tunes the network parameters by softmax classifier, and divides the input image into one of a large number of label classes. Secondly, according to the idea that the distance between image features extracted from same labels is close to each other, while the distance between the image features from different labels is away from each other, the algorithm fine-tunes network parameters by minimizing the verification loss function. Finally, the algorithm applies the high-level image features to train another classifier. According to this method, we achieve a network with better classification effect.(2)The algorithm applies hash codes to high-level semantic features extracted from original images, the image retrieval is achieved by comparing hamming distance between hash codes.Finally, a large number of experiments are carried out on CIFAR-10 image database and MINIST database. Comparing the classification accuracy and the average accuracy between various algorithms, the result indicates that the method proposed by this paper has better effect on classification and retrieval, and this method has certain reference significance and practical value for future study.