用于空中红外目标检测的增强单发多框检测器方法 
下载： 1189次
Enhancement of Single Shot Multibox Detector for Aerial Infrared Target Detection
1 中国科学院上海技术物理研究所, 上海 200083
2 中国科学院大学, 北京 100049
3 中国科学院红外探测与成像技术重点实验室, 上海 200083
图 & 表
图 1. SSD网络结构框图
Fig. 1. Structure of SSD network
下载图片 查看原文
图 2. 多种特征融合方法示意图。(a)池化;(b)转置卷积;(c)双向融合
Fig. 2. Schematics of multiple feature fusion methods. (a) Pooling; (b) transposed deconvolution; (c) bi-direction fusion
下载图片 查看原文
图 3. 语义分割分支结构图
Fig. 3. Diagram of semantic segmentation branch
下载图片 查看原文
图 4. 原始SSD与改进SSD检测小目标的结果对比。(a)原始SSD; (b)改进模型
Fig. 4. Comparison of detection results of small targets obtained by original SSD and improved SSD. (a) Original SSD; (b) improved model
下载图片 查看原文
图 5. 空中红外目标检测结果
Fig. 5. Detection results of infrared aerial targets
下载图片 查看原文
图 6. 空中红外目标召回率-准确率曲线对比
Fig. 6. Comparison of recall-precision curve of infrared aerial targets
下载图片 查看原文
表 1SSD_300×300卷积感受野、默认框映射图像区域
Table1. Convolution receptive field and mapping image region of default boxes of SSD_300×300
| Convolutional layer | Convolutional receptive field /(pixel×pixel) | Output scale of feature layer /(pixel×pixel) | Default boxes ratio | Mapping region scale /(pixel×pixel) |
|---|
| conv4_3 | 92×92 | 38×38 | 0.10 | 30×30 | | conv7 | 276×276 | 19×19 | 0.20 | 60×60 | | conv8_2 | 340×340 | 10×10 | 0.37 | 111×111 | | conv9_2 | 468×468 | 5×5 | 0.54 | 162×162 | | conv10_2 | 724×724 | 3×3 | 0.71 | 213×213 | | conv11_2 | 980×980 | 1×1 | 0.88 | 264×264 |
|
查看原文
表 2VOC2007数据集小目标检测结果
Table2. Small object detection results of VOC2007 dataset
| Method | mAP | Detection result |
|---|
| Aero plane | Bird | Boat | Bottle | Car | Dog | Sheep | Person |
|---|
| SSD_300×300 | 0.537 | 0.601 | 0.525 | 0.426 | 0.374 | 0.720 | 0.556 | 0.538 | 0.563 | | Proposed method | 0.608 | 0.685 | 0.570 | 0.534 | 0.503 | 0.748 | 0.597 | 0.652 | 0.591 |
|
查看原文
表 3红外数据集空中目标检测结果
Table3. Aerial target detection results of infrared dataset
| Method | mAP | Detection result |
|---|
| Fighter_J | Helicopter | Fighter_S | Airliner | Bird |
|---|
| SSD_300×300 | 0.618 | 0.659 | 0.615 | 0.266 | 0.819 | 0.732 | | YOLOv3-320 | 0.641 | 0.730 | 0.548 | 0.314 | 0.867 | 0.747 | | Proposed method | 0.705 | 0.784 | 0.636 | 0.485 | 0.822 | 0.796 |
|
查看原文
表 4每个分类网络的预测框数目及运行速度对比
Table4. Number of predictive boxes for each classi?ed network and speed comparison
| Method | Number of boxes | Total boxes | FPS |
|---|
| 75×75 | 38×38 | 19×19 | 10×10 | 5×5 | 3×3 | 1×1 |
|---|
| Original SSD | 0 | 4 | 6 | 6 | 6 | 4 | 4 | 8732 | 25.2 | | Modified SSD | 4 | 4 | 4 | 4 | 4 | 4 | 4 | 30260 | 9.4 |
|
查看原文
谢江荣, 李范鸣, 卫红, 李冰, 邵保泰. 用于空中红外目标检测的增强单发多框检测器方法[J]. 光学学报, 2019, 39(6): 0615001. Jiangrong Xie, Fanming Li, Hong Wei, Bing Li, Baotai Shao. Enhancement of Single Shot Multibox Detector for Aerial Infrared Target Detection[J]. Acta Optica Sinica, 2019, 39(6): 0615001.
PDF全文
