一种基于多尺度特征融合的目标检测算法 下载: 1422次
Multiscale Feature Fusion-Based Object Detection Algorithm
天津大学电气自动化与信息工程学院, 天津 300072
图 & 表
图 1. 基于多尺度特征融合的网络结构
Fig. 1. Network structure based on multi-scale feature fusion
下载图片 查看原文
图 2. 不同特征融合模块的网络结构。(a)FPN模块;(b)LFF模块;(c)HFF模块
Fig. 2. Network structure of different feature fusion modules. (a) FPN module; (b) LFF module; (c) HFF module
下载图片 查看原文
图 3. 加入特征融合模块前后的损失函数曲线。(a)加入前;(b)加入后
Fig. 3. Loss function curves before and after adding feature fusion module. (a) Before adding; (b) after adding
下载图片 查看原文
图 4. 加入特征融合模块前后的检测结果。(a)加入前;(b)加入后
Fig. 4. Detection results before and after adding feature fusion module. (a) Before adding; (b) after adding
下载图片 查看原文
表 1不同检测算法在PASCAL VOC数据集上的mAP值
Table1. mAP values of different detection algorithms on PASCAL VOC dataset
Algorithm | Backbone network | mAP /% |
---|
SSD300 | VGG-16 | 69.7 | RetinaNet | ResNet-50 | 70.2 | Libra RetinaNet | ResNet-50 | 70.4 | Proposed algorithm | ResNet-50 | 72.6 |
|
查看原文
表 2不同类别的AP值
Table2. AP values of different categories unit: %
Category | SSD300 | RetinaNet | Libra RetinaNet | Proposed algorithm |
---|
Aero | 75.7 | 75.4 | 75.4 | 77.4 | Bike | 78.4 | 80.2 | 79.6 | 80.1 | Bird | 67.2 | 72.1 | 72.1 | 73.5 | Boat | 64.4 | 60.7 | 64.8 | 67.3 | Bottle | 38.9 | 39.4 | 37.9 | 43.1 | Bus | 79.9 | 78.9 | 79.4 | 81.1 | Car | 83.3 | 79.0 | 79.0 | 80.3 | Cat | 84.6 | 86.4 | 85.1 | 87.4 | Chair | 49.5 | 52.5 | 52.6 | 54.9 | Cow | 67.1 | 64.1 | 68.0 | 74.1 | Category | SSD300 | RetinaNet | Libra RetinaNet | Proposed algorithm | Table | 63.7 | 67.7 | 66.5 | 68.5 | Dog | 79.2 | 80.1 | 80.5 | 82.2 | Horse | 80.5 | 79.4 | 78.9 | 81.4 | Mbike | 79.7 | 78.2 | 78.1 | 79.7 | Person | 75.2 | 74.2 | 74.5 | 75.6 | Plant | 37.2 | 43.0 | 42.3 | 45.3 | Sheep | 69.9 | 68.4 | 70.4 | 70.9 | Sofa | 68.5 | 71.3 | 71.3 | 72.0 | Train | 81.6 | 83.1 | 83.7 | 84.1 | TV | 69.3 | 70.3 | 68.8 | 72.7 |
|
查看原文
表 3消融学习的结果
Table3. Results of ablation study
HFF module | LFF module | mAP /% |
---|
| | 70.4 | √ | | 72.0 | | √ | 72.2 | √ | √ | 72.6 |
|
查看原文
表 4消融学习后每个类别的AP值
Table4. AP values of each category after ablation learning unit: %
Category | LFF module | HFF module |
---|
Aero | 76.5 | 77.9 | Bike | 80.2 | 80.5 | Bird | 74.1 | 72.6 | Boat | 66.0 | 66.6 | Bottle | 40.7 | 41.9 | Bus | 78.7 | 82.4 | Car | 80.2 | 79.4 | Cat | 86.1 | 86.8 | Chair | 54.9 | 53.2 | Cow | 70.9 | 69.2 | Table | 70.2 | 68.3 | Dog | 82.2 | 82.2 | Horse | 82.5 | 82.5 | Category | LFF module | HFF module | Mbike | 79.2 | 80.5 | Person | 75.0 | 75.0 | Plant | 44.7 | 45.0 | Sheep | 69.8 | 70.7 | Sofa | 73.5 | 73.3 | Train | 83.9 | 83.4 | TV | 70.5 | 71.6 |
|
查看原文
表 5不同检测算法在MSCOCO数据集上的AP值
Table5. AP values of different detection algorithms on MSCOCO dataset unit: %
Algorithm | Backbone | AP | AP50 | AP75 | APS | APM | APL |
---|
SSD512 | VGG-16 | 25.7 | 44.1 | 26.6 | 9.2 | 29.0 | 39.0 | RetinaNet | ResNet-50 | 35.6 | 55.6 | 38.1 | 20.8 | 39.5 | 46.1 | Libra RetinaNet | ResNet-50 | 37.5 | 56.9 | 39.9 | 22.4 | 41.4 | 49.2 | Proposed algorithm | ResNet-50 | 38.8 | 58.5 | 41.3 | 22.9 | 42.6 | 50.4 | RetinaNet | ResNet-101 | 37.8 | 57.5 | 40.8 | 20.9 | 42.1 | 49.6 | Libra RetinaNet | ResNet-101 | 39.1 | 58.6 | 41.7 | 22.6 | 43.8 | 51.4 | Proposed algorithm | ResNet-101 | 40.3 | 59.9 | 42.9 | 23.1 | 44.8 | 53.3 |
|
查看原文
表 6在COCO val-2017上消融学习后的结果
Table6. Results after ablation learning on COCO val-2017 unit: %
HFF module | LFF module | AP | AP50 | AP75 | APS | APM | APL |
---|
| | 37.5 | 56.9 | 39.9 | 22.4 | 41.4 | 49.2 | √ | | 38.5 | 58.3 | 41.0 | 22.4 | 42.6 | 50.2 | | √ | 38.5 | 58.4 | 40.9 | 23.8 | 42.4 | 49.7 | √ | √ | 38.8 | 58.5 | 41.3 | 22.9 | 42.6 | 50.4 |
|
查看原文
张涛, 张乐. 一种基于多尺度特征融合的目标检测算法[J]. 激光与光电子学进展, 2021, 58(2): 0215003. Tao Zhang, Le Zhang. Multiscale Feature Fusion-Based Object Detection Algorithm[J]. Laser & Optoelectronics Progress, 2021, 58(2): 0215003.