Research And Application Of Large-scale Tourist Attractions Image Retrieval Method Based On Deep Learning

Content-based image retrieval of tourist attractions is a basic task in the field of intelligence tourism.Image retrieval of tourist attractions is more challenging than other ordinary content-based image retrieval(CBIR)tasks because of the scenic spot images are photographed in different seasons,lighting conditions and angles.Hash retrieval is a commonly used image retrieval method in CBIR,and the hash encoding used in the algorithm has unique advantages in computing and storage.However,the traditional hash algorithm has the defect of poor feature expression and high coding redundancy,which caused unsatisfactory performance in the task of image retrieval of tourist attractions.In order to solve this problem,this paper uses a new hash code generation strategy based on deep learning that can end-to-end training hash maps.In order to overcome the shortcoming of the existing deep hash algorithms,some improvements have been made to the existing algorithms to improve the stability of the training and the retrieval performance,so as to achieve the fast and accurate retrieval of large-scale tourist attractions image data.Specifically,this article mainly includes the following aspects:(1)Aiming at the limitation of manual features in feature expression ability and generalization ability,the convolutional neural network is used in this paper as feature extraction sub-network to extract image features with better expression capabilities.Moreover,the network parameters are migrated by migration learning to further optimize network performance.(2)Comprehensive analysis of the merits and demerits of existing deep hash retrieval algorithms,the existing methods are improved in this Paper.First,block coding is used instead of the traditional fully connected layer to construct the hash layer to reduce the redundancy in the generated hash code.Second,the triplet loss function is employed as an objective function to optimize the network so that the sample can maintain similarity in the mapping process.In view of the problems existing in the sample selection of classical triplet loss function,an improved triplet sample selection strategy is used to ensure the stability of training and enhance the training effect of the network.(3)Using keywords to crawl pictures from the network to construct a Chinese popular tourist attractions image data set,which is utilized to verify the effectiveness ofthis method.And a simple prototype system has been implemented to demonstrate the retrieval effect of the algorithm in practical applications.By comparing various advanced deep hash algorithms with traditional hash algorithms on diverse evaluation criteria,the methods presented in this paper have certain advantages.The proposed method not only can train the depth hash network more quickly and steadily,but also can map the image data into binary hash code on the premise of maintaining the similarity relationship,which provides a new research idea for fast and accurate image retrieval of tourist attractions.