|
In image retrieval tasks, two main indicators are focused on: accuracy and efficiency. In general, the more the number of features, the higher the retrieval accuracy but the lower efficiency. Focusing on ensuring retrieval efficiency while improving retrieval accuracy, this paper designs an unsupervised image retrieval system that integrates multiple features. On UKbench, Holidays, UC Merced Land Use, and RSSCN7, the method proposed was validated. On UKbench, compared to methods based on shuffleNetV2, VGG-16 and Resnet18, the method proposed improves N-S by 0.1882, 0.2499, and 0.1985 respectively, while saving retrieval time by 0.0115s. On Holidays, compared to methods based on shuffleNetV2, VGG-16 and Resnet18, the method proposed improves MAP by 7.11%, 9.05%, and 7.27% respectively, while saving retrieval time by 0.0021s. On UC Merced Land Use, compared to methods based on shuffleNetV2, VGG16 and Resnet18, the method proposed improves accuracy by an average of 3.31%, 11.84%, and 6.29% respectively, while saving retrieval time by 0.0007s. On RSSCN7, compared to methods based on shuffleNetV2, VGG-16 and Resnet18, the method proposed improves accuracy by an average of 2.19%, 9.19%, and 7.18% respectively, while saving retrieval time by 0.0002s. The experimental results show that the proposed method can effectively improve retrieval accuracy and to some extent improve retrieval efficiency
|
