Acta Photonica Sinica, Volume. 49, Issue 1, 0128002(2020)
Feature Enhancement SSD Algorithm and Its Application in Remote Sensing Images Target Detection
Figures & Tables(11)
Fig. 2. FESSD算法框架图、浅层特征增强模块和深层特征增强模块图FESSD overall framework, shallow feature enhancement module and deep feature enhancement module
Fig. 3. 特征增强前后特征图对比Comparison of feature maps before and after feature enhancement
Fig. 5. 不同IOU情况下平均精度与召回率关系Recall vs. average of each class precision graph in different IOU conditions
Fig. 6. FESSD算法对光学遥感图像检测示例Examples of FESSD algorithm for optical remote sensing image detection
Fig. 7. Examples of FESSD and SSD algorithm migration experiment detection
Table 1. Self-collected data set statistics
View tableView in ArticleTable 1. Self-collected data set statistics
Data set Class #Image #Instances Target amount(percentage) Small Medium Large Training set Aircraft 320 2 382 1 491 890 1 Oiltank 220 1 735 884 845 6 Ship 400 1 317 659 588 70 Overpass 140 143 0 0 143 Playground 140 154 0 8 146 Test set Aircraft 101 970 571 398 1 Oiltank 78 759 374 385 0 Ship 70 415 210 195 10 Overpass 36 40 0 7 33 Playground 49 54 0 5 49
Table 2. Detection accuracy and speed of objects in remote sensing images by various algorithms
View tableView in ArticleTable 2. Detection accuracy and speed of objects in remote sensing images by various algorithms
Method Backbone Metric/% FPS Aircraft Oiltank Ship Over-pass Play-ground mAP(IOU=0.5) Faster R-CNN[ 5 ]VGG16 69.77 69.59 66.88 71.10 89.94 73.46 4.4 YOLO[ 6 ]GoogleNet 60.26 58.47 59.14 62.21 87.57 65.53 50.6 YOLO V2[ 9 ]DarkNet19 70.12 71.24 67.25 69.41 87.50 73.10 25.3 SSD300[ 7 ]VGG16 69.91 70.73 66.39 70.90 91.87 73.96 50.4 DSSD321[ 11 ]ResNet-101 75.75 73.11 70.32 71.22 90.52 76.18 4.3 ESSD321[ 12 ]VGG16 77.26 73.26 71.06 70.85 90.87 76.66 24.6 FFSSD300[ 13 ]VGG16 75.94 73.42 70.56 70.96 90.47 76.27 28.7 FESSD(ours)300 VGG16 79.01 76.12 72.27 71.05 91.96 78.08 48.6 SSD512[ 7 ]VGG16 71.56 72.28 70.04 71.29 92.05 75.44 28.6 DSSD513[ 11 ]ResNet-101 76.48 75.71 71.62 72.05 91.25 77.42 2.1 ESSD321[ 12 ]VGG16 77.45 75.89 72.34 71.44 91.66 77.76 12.3 FFSSD512[ 13 ]VGG16 77.23 76.89 72.36 71.59 91.61 77.94 13.4 FESSD(ours)512 VGG16 80.96 78.22 73.56 71.89 92.16 79.36 26.3
Table 3. The influence of feature enhancement on target detection in optical remote sensing images
View tableView in ArticleTable 3. The influence of feature enhancement on target detection in optical remote sensing images
SSD SFE DFE Metric/% FPS Aircraft Oiltank Ship Overpass Playground mAP √ 69.91 70.73 66.39 70.90 91.87 73.96 50.4 √ √ 78.38 75.45 71.63 70.06 90.91 77.29 35.6 √ √ 72.26 71.62 70.33 71.51 91.47 75.44 55.8 √ √ √ 79.01 76.12 72.27 71.05 91.96 78.08 48.6
Table 4. The influence of each independent module on target detection in optical remote sensing images
View tableView in ArticleTable 4. The influence of each independent module on target detection in optical remote sensing images
mAP +DFE_11 +DFE_10 +DFE_9 +DFE_8 +SFE2 +SFE1 73.96 76.57 √ 77.29 √ √ 77.45 √ √ √ 77.86 √ √ √ √ 77.99 √ √ √ √ √ 78.08 √ √ √ √ √ √
Tools
Get Citation
Copy Citation Text
Wen-xu SHI, Dai-lun TAN, Sheng-li BAO. Feature Enhancement SSD Algorithm and Its Application in Remote Sensing Images Target Detection[J]. Acta Photonica Sinica, 2020, 49(1): 0128002
Paper Information
Received: Jul. 29, 2019
Accepted: Oct. 22, 2019
Published Online: Mar. 19, 2020
The Author Email: BAO Sheng-li (baohigh@casit.com.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20