Deep Cross-Dimensional Attention Hashing for Image Retrieval

Nowadays, people’s lives are filled with a huge amount of picture information, and image retrieval tasks widely needed. Deep hashing methods extensively used to manage such demands due their rate memory consumption. The problem conventional deep techniques, however, is that high dimensional semantic content in the cannot be effectively articulated insufficient unbalanced feature extraction. This paper offers cross-dimensional attention (DCDAH) method considering flaws extraction, important points this as follows. proposes (CDA) module embedded ResNet18; can capture cross-dimension interaction maps calculate weight because its special branch. For map acquired by convolutional neural network (CNN), each branch takes different rotation measurements residual transformations process it. To prevent DCDAH model from becoming too complex, CDA designed have characteristics low computational overhead. introduces scheme reduce dimension tensors, which computation retain abundant representation. map, Maxpool Avgpool performed, respectively, two results connected. significantly enhances performance, according studies on CIFAR10 NUS-WIDE data sets.