An unsupervised hash retrieval algorithm based on feature co-occurrence

The existing algorithms of unsupervised hash retrieval focus on the information loss in the process of hash mapping and the quality of hash generation, but ignore the impact of image features on the retrieval accuracy. In order to further improve the retrieval accuracy, this paper proposes an improved Unsupervised Hash retrieval algorithm based on Feature Co-occurrence (UHFC), which is divided into two stages: deep feature extraction and unsupervised hash generation. In order to improve the quality of image features, UHFC introduces a co-occurrence layer after the last convolution layer of Convolutional Neural Network (CNN) structure to extract the dependency relationship between features. The mean value of co-occurrence activation value is used to represent the degree of co-occurrence to solve the problem of inconsistent co-present value of the same two channels in the original co-occurrence operation. Then, in the feature fusion part of UHFC, an Attentional Feature Fusion method based on Spatial attention(AFF-S) mechanism is designed for co-occurrence feature fusion. By self-learning the weight of co-occurrence feature and depth feature fusion by attention mechanism, the interference of background factors in the process of feature fusion is reduced, and the expressive ability of final image features is improved. Finally, according to the optimal transmission strategy, UHFC adopts Bi-half distributed hash coding to supervise the mapping process of image features to hash code, and adds a classification layer after the hash layer to further improve the image information contained in the hash code through KL loss. In the whole training process, no data set labeling is required to realize the generation of unsupervised hash. Experiments have shown that UHFC better improve quality of hash code, in Flickr25k and Nus - wide data sets its mean Average Precision (mAP) reached 87.8% and 82.8% respectively, compared to the baseline method is increased by 2.1% and 1.2%, respectively, effect is obvious.