基于聚类式区域生成的全卷积目标检测网络 
下载： 813次
Full-Convolution Object Detection Network Based on Clustering Region Generation
江南大学物联网工程学院, 轻工过程先进控制教育部重点实验室, 江苏 无锡 214122
图 & 表
图 1. R-FCN结构图
Fig. 1. Structure of R-FCN
下载图片 查看原文
图 2. RPN结构图
Fig. 2. Structure of RPN
下载图片 查看原文
图 3. R-FCN检测结果。 (a)定位不精确;(b)检测误差严重
Fig. 3. Detection results of R-FCN. (a) Inaccurate locating; (b) serious detection error
下载图片 查看原文
图 4. 改进后的网络框架
Fig. 4. Improved network frame
下载图片 查看原文
图 5. RPN聚类网络基本结构
Fig. 5. Basic structure of RPN clustering network
下载图片 查看原文
图 6. 聚类检测结果。(a)不同K值下的平均IOU;(b)不同K值下的检测精度;(c)不同K值下聚类耗费时间
Fig. 6. Clustering detection results. (a) Average IOU with different K values; (b) detection accuracy with different K values; (c) clustering consuming with different K values
下载图片 查看原文
图 7. 算法改进前后检测效果对比图。(a) R-FCN检测结果;(b)改进算法检测结果
Fig. 7. Comparison of detection results before and after algorithm improvement. (a) Detection results of R-FCN; (b) detection results of improved algorithm
下载图片 查看原文
表 1ResNet-50下,不同方法的检测结果比较
Table1. Detection results with different methods based on ResNet-50
| Backbonenetwork | Method | mAP /% | Detectiontime /s |
|---|
| ResNet-50 | Faster R-CNN | 76.60 | 0.420 | | R-FCN | N/A | N/A | | Proposed | 79.04 | 0.031 | | Faster R-CNN (OHEM) | N/A | N/A | | R-FCN (OHEM) | 77.40 | 0.099 | | Proposed (OHEM) | 83.36 | 0.031 |
|
查看原文
表 2ResNet-101下,不同方法的检测结果比较
Table2. Detection results with different methods based on RseNet-101
| Backbonenetwork | Method | mAP /% | Detectiontime /s |
|---|
| ResNet-101 | Faster R-CNN | 76.40 | 0.420 | | R-FCN | 76.60 | 0.170 | | Proposed | 81.01 | 0.046 | | Faster R-CNN (OHEM) | 79.44 | 0.042 | | R-FCN (OHEM) | 79.50 | 0.170 | | Proposed (OHEM) | 84.64 | 0.046 |
|
查看原文
表 3ResNet-101下不同方法的各类检测结果
Table3. All kinds of detection results with different methods based on RseNet-101
| Method | mAP/% | Areo | Cat | Bird | Boat | Bottle | Bus | Plant | Bike | Chair | Cow |
|---|
| R-FCN | 79.5 | 82.5 | 88.4 | 83.7 | 69.0 | 69.2 | 87.5 | 54.1 | 83.7 | 65.4 | 87.3 | | Ours | 81.01 | 82.0 | 90.8 | 82.8 | 79.3 | 59.2 | 89.4 | 58.3 | 82.8 | 59.6 | 88.9 | | Proposed (OHEM) | 84.64 | 82.5 | 90.7 | 88.5 | 81.3 | 71.4 | 89.9 | 66.7 | 88.5 | 72.3 | 89.7 | | Method | mAP/% | Table | Dog | Horse | Bike | Person | Car | Sheep | Sofa | Train | TV | | R-FCN | 79.5 | 72.1 | 87.9 | 88.3 | 81.3 | 79.8 | 88.4 | 79.6 | 78.8 | 87.1 | 79.5 | | Ours | 81.01 | 75.1 | 90.8 | 89.9 | 84.2 | 79.2 | 85.5 | 86.3 | 84.3 | 90.2 | 77.9 | | Proposed (OHEM) | 84.64 | 81.1 | 90.6 | 90.3 | 88.1 | 80.3 | 88.3 | 89.2 | 85.1 | 90.4 | 86.0 |
|
查看原文
表 4K值普适性实验结果
Table4. Experimental results of generalization of K value
| Backbonenetwork | Method | mAP /% | Detectiontime /s |
|---|
| ResNet-101 | R-FCN (OHEM) | 40.06 | 0.170 | | Proposed (OHEM) | 41.16 | 0.046 |
|
查看原文
潘志浩, 陈莹. 基于聚类式区域生成的全卷积目标检测网络[J]. 激光与光电子学进展, 2019, 56(15): 151001. Zhihao Pan, Ying Chen. Full-Convolution Object Detection Network Based on Clustering Region Generation[J]. Laser & Optoelectronics Progress, 2019, 56(15): 151001.
PDF全文
