Enhanced supervised image retrieval method based on image subjects

With the increase of data size, hash retrieval has received widespread attention due to its efficiency and scalability. However, existing deep hashing methods usually extract features based on the whole image, which is susceptible to redundant information and limits the retrieval performance. In addition, these methods lack effective modelling of intra-class features and are difficult to adapt to complex feature changes. To address the above problems, a novel deep hashing method is proposed, combining image subject feature extraction, implicit feature enhancement and multi-error co-optimisation strategies. Firstly, the feature extraction strategy focusing on the image subject content is adopted to eliminate redundant information; secondly, the implicit feature enhancement module is used to dynamically model the category distribution and enhance the intra-class semantic adaptation; finally, the loss distribution is optimised by combining the classification error and hashing error to enhance the model robustness. Experimental results show that the retrieval performance of this method on multiple public datasets is significantly better than existing algorithms.