一种改进的多门控特征金字塔网络 
下载： 1097次
An Improved Multi-Gate Feature Pyramid Network
火箭军工程大学导弹工程学院, 陕西 西安 710025
图 & 表
图 1. FPN结构
Fig. 1. FPN structure
下载图片 查看原文
图 2. 改进的FPN结构
Fig. 2. Improved FPN structure
下载图片 查看原文
图 3. 记忆和滤波通道的结构
Fig. 3. Memory and filter channel structure
下载图片 查看原文
图 4. MSSD网络结构
Fig. 4. MSSD network structure
下载图片 查看原文
图 5. SSD算法和MSSD算法的可视化对比。(a)(c)(e)(g) SSD算法;(b)(d)(f)(h) MSSD算法
Fig. 5. Visual comparison of SSD algorithm and MSSD algorithm. (a)(c)(e)(g) SSD algorithm; (b)(d)(f)(h) MSSD algorithm
下载图片 查看原文
表 1加入FPN结构的SSD网络与MSSD网络的性能对比
Table1. Comparison between SSD network with FPN structure and MSSD network
| Method | Feature layers | mAP /% | FPS |
|---|
| SSD[6] | Conv4-Pool11 | 77.5 | 46(Titan X) | | SSD+FPN | Conv4-Conv9 | 77.7 | 53.9 | | MSSD | Conv4-Pool11 | 78.8 | 27.1 | | MSSD | Conv4-Conv9 | 79.0 | 31.7 |
|
查看原文
表 2融合结构的改进效果对比
Table2. Comparison of improved effects of fusion structure
| Conv 1×1 | Feature fusion mode | Deconv or bilinear interpolation | mAP /% | FPS |
|---|
| √ | Sum | Deconv | 78.9 | 31.2 | | Sum | Deconv | 79.0 | 31.7 | | √ | Sum | Bilinear interpolation | 78.5 | 40.1 | | Sum | Bilinear interpolation | 78.5 | 40.2 | | Max | Deconv | 78.2 | 36.4 | | Max | Bilinear interpolation | 76.6 | 42.2 |
|
查看原文
表 3记忆和滤波通道的有效性分析
Table3. Analysis of effectiveness of memory and filter channels
| Number | Forget gate | Input gate | Output gate | mAP /% |
|---|
| 1 | √ | | | / | | 2 | | √ | | 75.2 | | 3 | | | √ | / | | 4 | √ | √ | | 78.6 | | 5 | √ | | √ | / | | 6 | | √ | √ | 77.9 | | 7 | √ | √ | √ | 79 |
|
查看原文
表 4改变基础网络性能对比
Table4. Performance comparison of different basic networks
| Method | Network | mAP /% | FPS |
|---|
| SSD[11] | ResNet101 | 77.1 | 18.9(Titan X) | | SSD+FPN | ResNet101 | 78.7 | 39.1 | | MSSD | ResNet101 | 79.3 | 23.7 | | MSSD | VGG16 | 79.0 | 31.7 |
|
查看原文
表 5各类先进深度学习算法在VOC2007数据集上精度对比
Table5. Accuracy comparison of various advanced deep learning algorithms on VOC2007 dataset
| Method | Network | mAP /% |
|---|
| Aero | Bike | Bird | Boat | Bottle | Bus | Car | Cat | Chair | Cow | Table | Dog | Horse | Mbike | Person | Plant | Sheep | Sofa | Train | Tv | Average |
|---|
| Faster[5] | VGG16 | 76.5 | 79.0 | 70.9 | 65.5 | 52.1 | 83.1 | 84.7 | 86.4 | 52.0 | 81.9 | 65.7 | 84.8 | 84.6 | 77.5 | 76.7 | 38.8 | 73.6 | 73.9 | 83.0 | 72.6 | 73.2 | | ION[23] | VGG16 | 79.2 | 83.1 | 77.6 | 65.6 | 54.9 | 85.4 | 85.1 | 87.0 | 54.4 | 80.6 | 73.8 | 85.3 | 82.2 | 82.2 | 74.4 | 47.1 | 75.8 | 72.7 | 84.2 | 80.4 | 75.6 | | Faster[18] | ResNet101 | 79.8 | 80.7 | 76.2 | 68.3 | 55.9 | 85.1 | 85.3 | 89.8 | 56.7 | 87.8 | 69.4 | 88.3 | 88.9 | 80.9 | 78.4 | 41.7 | 78.6 | 79.8 | 85.3 | 72.0 | 76.4 | | MR-CNN[24] | VGG16 | 80.3 | 84.1 | 78.5 | 70.8 | 68.5 | 88.0 | 85.9 | 87.8 | 60.3 | 85.2 | 73.7 | 87.2 | 86.5 | 85.0 | 76.4 | 48.5 | 76.3 | 75.5 | 85.0 | 81.0 | 78.2 | | R-FCN[22] | ResNet101 | 79.9 | 87.2 | 81.5 | 72.0 | 69.8 | 86.8 | 88.5 | 89.8 | 67.0 | 88.1 | 74.5 | 89.8 | 90.6 | 79.9 | 81.2 | 53.7 | 81.8 | 81.5 | 85.9 | 79.9 | 80.5 | | SSD300[6] | VGG16 | 79.5 | 83.9 | 76.0 | 69.6 | 50.5 | 87.0 | 85.7 | 88.1 | 60.3 | 81.5 | 77.0 | 86.1 | 87.5 | 83.97 | 79.4 | 52.3 | 77.9 | 79.5 | 87.6 | 76.8 | 77.5 | | SSD512[6] | VGG16 | 84.8 | 85.1 | 81.5 | 73.0 | 57.8 | 87.8 | 88.3 | 87.4 | 63.5 | 85.4 | 73.2 | 86.2 | 86.7 | 83.9 | 82.5 | 55.6 | 81.7 | 79.0 | 86.6 | 80.0 | 79.5 | | DSSD321[11] | ResNet101 | 81.9 | 84.9 | 80.5 | 68.4 | 53.9 | 85.6 | 86.2 | 88.9 | 61.1 | 83.5 | 78.7 | 86.7 | 88.7 | 86.7 | 79.7 | 51.7 | 78.0 | 80.9 | 87.2 | 79.4 | 78.6 | | DSSD513[11] | ResNet101 | 86.6 | 86.2 | 82.6 | 74.9 | 62.5 | 89.0 | 88.7 | 88.8 | 65.2 | 87.0 | 78.7 | 88.2 | 89.0 | 87.5 | 83.7 | 51.1 | 86.3 | 81.6 | 85.7 | 83.7 | 81.5 | | MDSSD300[25] | VGG16 | 86.5 | 87.6 | 78.9 | 70.6 | 55.0 | 86.9 | 87.0 | 88.1 | 58.5 | 84.8 | 73.4 | 84.8 | 89.2 | 88.1 | 78.0 | 52.3 | 78.6 | 74.5 | 86.8 | 80.7 | 78.6 | | MDSSD512[25] | VGG16 | 88.8 | 88.7 | 83.2 | 73.7 | 58.3 | 88.2 | 89.3 | 87.4 | 62.4 | 85.1 | 75.1 | 84.7 | 89.7 | 88.3 | 83.2 | 56.7 | 84.0 | 77.4 | 83.9 | 77.6 | 80.3 | | YOLOV3[8] | Darknet53 | 85.5 | 85.5 | 75.6 | 70.0 | 66.5 | 87.6 | 87.7 | 89.4 | 64.3 | 83.5 | 73.6 | 85.9 | 86.9 | 86.2 | 83.3 | 56.2 | 75.3 | 78.0 | 86.4 | 77.8 | 79.2 | | MSSD300 | ResNet101 | 81.2 | 87.2 | 78.7 | 72.7 | 53.4 | 86.4 | 85.6 | 89.1 | 63.1 | 84.5 | 80.0 | 87.5 | 88.9 | 84.8 | 78.8 | 54.5 | 80.9 | 83.2 | 87.1 | 77.4 | 79.3 | | MSSD300 | VGG16 | 81.6 | 85.8 | 78.0 | 74.0 | 55.3 | 86.2 | 86.5 | 88.2 | 64.6 | 85.9 | 76.9 | 85.4 | 87.7 | 85.2 | 79.5 | 51.1 | 78.8 | 80.8 | 87.9 | 78.6 | 79.0 | | MSSD500 | VGG16 | 87.6 | 87.2 | 83.7 | 75.5 | 57.8 | 86.7 | 88.4 | 89.5 | 66.3 | 84.6 | 78.9 | 86.8 | 88.1 | 86.4 | 83.3 | 58.0 | 81.4 | 81.0 | 88.4 | 78.1 | 81.0 |
|
查看原文
表 6各类基于SSD的改进算法的测试结果对比
Table6. Comparison of test results of various improved algorithms based on SSD
| Method | Network | FPS | GPU | #proposals | Input size /(pixel×pixel) | mAP /% |
|---|
| SSD300[6] | VGG16 | 46 | Titan X | 8732 | 300×300 | 77.2 | | SSD512[6] | VGG16 | 19 | Titan X | 24564 | 512×512 | 78.5 | | MDSSD300[25] | VGG16 | 38.5 | 1080Ti | 44530 | 300×300 | 78.6 | | MDSSD512[25] | VGG16 | 17.3 | 1080Ti | - | 512×512 | 80.3 | | DSSD321[11] | ResNet101 | 9.5 | Titan X | 17088 | 321×321 | 78.6 | | DSSD513[11] | ResNet101 | 5.5 | Titan X | 43688 | 513×513 | 81.5 | | RSSD300[14] | VGG16 | 35 | Titan X | 8732 | 300×300 | 78.5 | | RSSD512[14] | VGG16 | 16.6 | Titan X | 24564 | 512×512 | 80.8 | | MSSD300 | ResNet101 | 23.7 | 1080Ti | 8728 | 300×300 | 79.3 | | MSSD300 | VGG16 | 31.7 | 1080Ti | 8732 | 300×300 | 79.0 | | MSSD512 | VGG16 | 17.3 | 1080Ti | 24564 | 512×512 | 81.0 |
|
查看原文
表 7SSD算法和MSSD算法对小目标的检测精度
Table7. Detection accuracy of small targets by SSD algorithm and MSSD algorithm
| Method | mAP /% |
|---|
| SSD[6] | 55.4 | | MSSD | 56.3 |
|
查看原文
赵彤, 刘洁瑜, 沈强. 一种改进的多门控特征金字塔网络[J]. 光学学报, 2019, 39(8): 0815005. Tong Zhao, Jieyu Liu, Qiang Shen. An Improved Multi-Gate Feature Pyramid Network[J]. Acta Optica Sinica, 2019, 39(8): 0815005.
PDF全文
