基于最大间隔的半监督图像搜索重排序方法 
下载： 800次
A Max Margin Based Semi-Supervised Reranking Method
天津大学电气自动化与信息工程学院, 天津 300072
图 & 表
图 1. 提出的重排序算法流程图
Fig. 1. Flow chart of the proposed reranking algorithm
下载图片 查看原文
图 2. 基于查询词“angel”的(a)初始排序结果与(b)重排序结果对比
Fig. 2. Performance comparison between (a) initial search results and (b) reranking results based on query "angel"
下载图片 查看原文
图 3. 不同类别数据的实验结果比较
Fig. 3. Performance comparison of different datasets
下载图片 查看原文
表 1不同惩罚因子下的性能比较
Table1. Performance comparison for different trade-off parameters
| C | Depth |
|---|
| 10 | 20 | 30 | 40 | 50 | 60 | 70 | 80 | 90 | 100 |
|---|
| 0.01 | 0.634 | 0.590 | 0.572 | 0.563 | 0.559 | 0.557 | 0.555 | 0.555 | 0.558 | 0.562 | | 0.1 | 0.807 | 0.722 | 0.682 | 0.659 | 0.644 | 0.634 | 0.628 | 0.625 | 0.624 | 0.626 | | 1 | 0.862 | 0.766 | 0.723 | 0.697 | 0.679 | 0.667 | 0.659 | 0.656 | 0.653 | 0.653 | | 10 | 0.861 | 0.766 | 0.722 | 0.695 | 0.676 | 0.666 | 0.658 | 0.655 | 0.652 | 0.652 | | 100 | 0.861 | 0.766 | 0.722 | 0.695 | 0.676 | 0.666 | 0.658 | 0.654 | 0.652 | 0.652 |
|
查看原文
表 2每个相关性等级不同标注样本个数的性能比较
Table2. Performance comparison for different labeled numbers
| k | Depth |
|---|
| 10 | 20 | 30 | 40 | 50 | 60 | 70 | 80 | 90 | 100 |
|---|
| 5 | 0.862 | 0.766 | 0.723 | 0.697 | 0.679 | 0.667 | 0.659 | 0.656 | 0.653 | 0.653 | | 10 | 0.930 | 0.844 | 0.783 | 0.747 | 0.728 | 0.714 | 0.704 | 0.698 | 0.694 | 0.693 | | 15 | 0.866 | 0.879 | 0.827 | 0.789 | 0.762 | 0.744 | 0.731 | 0.723 | 0.717 | 0.715 | | 20 | 0.803 | 0.851 | 0.847 | 0.812 | 0.786 | 0.766 | 0.751 | 0.742 | 0.735 | 0.730 |
|
查看原文
表 3排序函数中设置的不同相关性等级的不同排序分数间隔性能比较
Table3. Performance comparison for different ranking fractional intervals
| m | Depth |
|---|
| 10 | 20 | 30 | 40 | 50 | 60 | 70 | 80 | 90 | 100 |
|---|
| 0 | 0.769 | 0.695 | 0.661 | 0.640 | 0.626 | 0.618 | 0.613 | 0.610 | 0.610 | 0.614 | | 0.5 | 0.862 | 0.766 | 0.723 | 0.697 | 0.679 | 0.667 | 0.659 | 0.656 | 0.653 | 0.653 | | 1 | 0.856 | 0.764 | 0.722 | 0.697 | 0.679 | 0.666 | 0.659 | 0.655 | 0.653 | 0.653 | | 1.5 | 0.850 | 0.761 | 0.719 | 0.692 | 0.677 | 0.665 | 0.656 | 0.652 | 0.650 | 0.650 | | 2 | 0.844 | 0.756 | 0.715 | 0.686 | 0.671 | 0.661 | 0.654 | 0.650 | 0.647 | 0.647 | | 2.5 | 0.838 | 0.752 | 0.711 | 0.683 | 0.666 | 0.657 | 0.650 | 0.647 | 0.644 | 0.643 | | 3 | 0.832 | 0.746 | 0.705 | 0.680 | 0.661 | 0.653 | 0.646 | 0.642 | 0.639 | 0.640 | | 3.5 | 0.827 | 0.740 | 0.699 | 0.674 | 0.657 | 0.648 | 0.641 | 0.637 | 0.635 | 0.637 | | 4 | 0.823 | 0.736 | 0.695 | 0.670 | 0.654 | 0.643 | 0.637 | 0.633 | 0.631 | 0.634 | | 4.5 | 0.815 | 0.729 | 0.688 | 0.663 | 0.648 | 0.638 | 0.632 | 0.628 | 0.627 | 0.630 | | 5 | 0.807 | 0.722 | 0.682 | 0.659 | 0.644 | 0.634 | 0.628 | 0.625 | 0.624 | 0.626 |
|
查看原文
表 4不同图像搜索重排序方法的实验结果比较
Table4. Performance comparison for different image search reranking methods
| Method | Depth |
|---|
| 10 | 20 | 30 | 40 | 50 | 60 | 70 | 80 | 90 | 100 |
|---|
| RankSVM | 0.670 | 0.665 | 0.659 | 0.649 | 0.641 | 0.636 | 0.634 | 0.634 | 0.633 | 0.636 | | RankSVM+LPP | 0.801 | 0.735 | 0.702 | 0.679 | 0.669 | 0.659 | 0.654 | 0.651 | 0.649 | 0.650 | | RANGE | 0.835 | 0.753 | 0.717 | 0.692 | 0.676 | 0.666 | 0.660 | 0.658 | 0.656 | 0.657 | | Proposed | 0.859 | 0.760 | 0.719 | 0.691 | 0.676 | 0.668 | 0.662 | 0.659 | 0.658 | 0.658 |
|
查看原文
张桐喆, 苏育挺, 郭洪斌. 基于最大间隔的半监督图像搜索重排序方法[J]. 激光与光电子学进展, 2018, 55(11): 111001. Tongzhe Zhang, Yuting Su, Hongbin Guo. A Max Margin Based Semi-Supervised Reranking Method[J]. Laser & Optoelectronics Progress, 2018, 55(11): 111001.
PDF全文
