Research On Large-scale Image Retrieval Method Based On Hash Learning

The rapid rise of the digital industry has resulted in the emergence of tens of thousands of multimedia data such as images,texts,and videos on the Internet.These multimedia data are often used in data retrieval,recognition,classification and other tasks.Hashing,as a widely studied Approximate Nearest Neighbor Search(ANNS)method,speeds up retrieval by mapping media data into compact binary codes while maintaining the similarity of the original data as much as possible,effectively reducing computing and storage overhead,and significantly saving resources.Therefore,it has attracted extensive attention from academia and industry due to its advantages such as low storage and high efficiency,and has become one of the important methods for large-scale image retrieval.In the task of large-scale image retrieval,some existing image hashing methods usually predefine a hash code of a specific length according to the task requirements during the learning process,and then use it as the target of model learning.When task requirements or conditions change,the model may need to be retrained multiple times to learn new length hash codes,which increases the cost of computing power and hardware resources,and greatly reduces the training efficiency.Secondly,the learning of a single hash code tends to ignore potential correlations among multiple hash codes.Specifically,the hash code is a reflection of the overall semantic information of the image data and its length is variable.An image sample can be represented by one or more lengths of hash codes,and hash codes of different lengths are expressions of the semantic information of the original image sample.Then there should be correlation and consistency between these hash codes.Moreover,the learning of single-length hash codes lacks this semantic relevance,which causes the lack of image-specific semantic information to a certain extent.It affects the quality of image sample representation,resulting in low image retrieval accuracy.In response to the above problems,this paper proposes a series of hashing methods.The main contributions of this paper are as follows:(1)We propose a deep hashing method based on multi-task learning also named as DMLH(Deep multiple length hashing via multi-task learning).For some existing methods,the specific length hash code learning causes high resource overhead for repeated model training.Meanwhile,they ignore the potential correlation among these multiple hash codes etc.To solve the above issues,this paper develops a novel deep hashing framework.Specifically,it constructs a multi-branch deep network based on hard parameter sharing to simultaneously learn hash codes of various lengths to meet various task requirements,effectively solving the problem of the model’s single-length hash code repeated training limitations.Besides,it uses pairwise similarity loss and mutual information loss to maintain the similarity of the original sample space and mine potential associations between multiple hash codes,which promises the quality of image hash codes,further enhancing the retrieval performance.(2)We propose an Attention-oriented deep multi-task hash learning method named as ADMTH.Aiming at the problem that the backbone network of some existing deep hashing methods does not learn enough image discriminative information.Building on the first work,this work combines the channel attention mechanism to weight the importance of feature maps from the channel level,making the learned image features more discriminative.In addition,this work also conducts a more in-depth exploration of multi-task learning,introducing multi-task learning training efficiency analysis,algorithm time complexity analysis,multi-task learning method split analysis,etc.The better hash code representation is obtained,which improves the accuracy of model retrieval.(3)We propose a deep hashing optimization method based on adaptive pairwise similarity matrix learning.Some existing deep hashing methods use predefined pairwise similarity matrices as supervisory information and ignore the shortcomings of detailed similarity relationships between raw data categories.Based on the two studies mentioned above,this work proposes an improved pairwise similarity matrix optimization method,which initializes the fixed pairwise similarity matrix as the parameters learned by the network,and dynamically learns the relationship between image categories as the network iteratively updates.It makes the learned hash code more accurately reflect the similarity relationship between the original samples,thereby improving the quality of the hash code representation and promoting the image retrieval accuracy.