Acta Photonica Sinica, Volume. 52, Issue 1, 0110003(2023)
Enhancing Remote Sensing Image Unsupervised Hashing Cross-modal Correlation with Similarity Matrix
Figures & Tables(10)
Fig. 2. Remote sensing image and corresponding captions from the datasets
Fig. 5. Model performance under different parameter settings（mAP@20）
Table 1. The mAP@20 comparison of different methods on the RSICD dataset
View tableView in ArticleTable 1. The mAP@20 comparison of different methods on the RSICD dataset
Method Image to Text（I→T） Text to Image（T→I） B=16 B=32 B=64 B=128 B=16 B=32 B=64 B=128 CPAH 0.428 0.587 0.636 0.696 0.452 0.598 0.667 0.706 DJSRH 0.411 0.665 0.688 0.722 0.422 0.685 0.705 0.733 JDSH 0.385 0.720 0.796 0.815 0.418 0.751 0.799 0.815 DUCH 0.684 0.791 0.836 0.829 0.697 0.780 0.824 0.826 Proposed 0.708 0.802 0.823 0.832 0.736 0.818 0.845 0.850
Table 2. The mAP@20 comparison of different methods on the UCM dataset
View tableView in ArticleTable 2. The mAP@20 comparison of different methods on the UCM dataset
Method Image to Text（I→T） Text to Image（T→I） B=16 B=32 B=64 B=128 B=16 B=32 B=64 B=128 CPAH 0.706 0.802 0.891 0.914 0.782 0.891 0.987 0.982 DJSRH 0.686 0.711 0.735 0.754 0.738 0.755 0.776 0.800 JDSH 0.462 0.751 0.82 0.829 0.509 0.794 0.884 0.904 DUCH 0.760 0.794 0.844 0.870 0.799 0.851 0.916 0.927 Proposed 0.789 0.816 0.841 0.860 0.848 0.894 0.926 0.951
Table 3. The mAP@20 comparison of different methods after optimization
View tableView in ArticleTable 3. The mAP@20 comparison of different methods after optimization
Dataset Method Image to Text（I→T） Text to Image（T→I） B=16 B=32 B=64 B=128 B=16 B=32 B=64 B=128 RSICD DUCH 0.698 0.795 0.838 0.838 0.732 0.825 0.85 0.856 Proposed 0.748 0.815 0.839 0.85 0.796 0.845 0.868 0.872 Proposed（vit） 0.805 0.875 0.893 0.907 0.819 0.882 0.892 0.891 UCM DUCH 0.768 0.824 0.896 0.910 0.802 0.862 0.938 0.953 Proposed 0.831 0.871 0.899 0.912 0.873 0.922 0.942 0.957 Proposed（vit） 0.905 0.915 0.933 0.939 0.923 0.947 0.961 0.969
Table 4. Ablation experiment results（50%training）
View tableView in ArticleTable 4. Ablation experiment results（50%training）
Dataset Method Image to Text（I→T） Text to Image（T→I） B=16 B=32 B=64 B=128 B=16 B=32 B=64 B=128 RSICD Lm+Lc 0.708 0.802 0.823 0.832 0.736 0.818 0.845 0.850 Lm 0.632 0.770 0.817 0.834 0.656 0.785 0.824 0.848 Lc 0.703 0.787 0.818 0.828 0.710 0.796 0.820 0.833 UCM Lm+Lc 0.789 0.816 0.841 0.86 0.848 0.894 0.926 0.951 Lm 0.747 0.817 0.851 0.859 0.813 0.877 0.934 0.943 Lc 0.760 0.808 0.842 0.863 0.823 0.890 0.935 0.948
Table 5. Ablation experiment results（70%training）
View tableView in ArticleTable 5. Ablation experiment results（70%training）
Dataset Method Image to Text（I→T） Text to Image（T→I） B=16 B=32 B=64 B=128 B=16 B=32 B=64 B=128 RSICD Lm+Lc 0.748 0.815 0.839 0.85 0.796 0.845 0.868 0.872 Lm 0.682 0.802 0.842 0.851 0.692 0.820 0.862 0.869 Lc 0.741 0.812 0.836 0.849 0.770 0.828 0.851 0.867 UCM Lm+Lc 0.831 0.871 0.899 0.912 0.873 0.922 0.942 0.957 Lm 0.817 0.859 0.869 0.879 0.841 0.898 0.913 0.930 Lc 0.835 0.867 0.885 0.898 0.869 0.917 0.940 0.952
Tools
Get Citation
Copy Citation Text
Haoran LI, Wei XIONG, Yaqi CUI, Xiangqi GU, Pingliang XU. Enhancing Remote Sensing Image Unsupervised Hashing Cross-modal Correlation with Similarity Matrix[J]. Acta Photonica Sinica, 2023, 52(1): 0110003
Paper Information
Category:
Received: Jul. 6, 2022
Accepted: Aug. 18, 2022
Published Online: Feb. 27, 2023
The Author Email: XIONG Wei (xiongwei@csif.org.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20