基于可變形卷積神經(jīng)網(wǎng)絡(luò)的遙感影像密集區(qū)域車輛檢測方法
doi: 10.11999/JEIT180209
-
1.
中國科學(xué)院電子學(xué)研究所空間信息處理與應(yīng)用系統(tǒng)技術(shù)重點(diǎn)實(shí)驗(yàn)室 ??北京 ??100190
-
2.
中國科學(xué)院大學(xué) ??北京 ??100049
-
3.
中央蘭開夏大學(xué) ??英國 ??PR1 2HE
Vehicle Detection in Remote Sensing Images of Dense Areas Based on Deformable Convolution Neural Network
-
1.
Key Laboratory of Technology in Geo-spatial Information Processing and Application System, Institute of Electronics, Chinese Academy of Sciences, Beijing 100190, China
-
2.
University of Chinese Academy of Sciences, Beijing 100049, China
-
3.
University of Central Lancashire (UCLan), Preston PR1 2HE, United Kingdom
-
摘要: 車輛檢測是遙感圖像分析領(lǐng)域的熱點(diǎn)研究內(nèi)容之一,車輛目標(biāo)的智能提取和識別,對于交通管理、城市建設(shè)有重要意義。在遙感領(lǐng)域中,現(xiàn)有基于卷積神經(jīng)網(wǎng)絡(luò)的車輛檢測方法存在實(shí)現(xiàn)過程復(fù)雜并且對于車輛密集區(qū)域檢測效果不理想的缺陷。針對上述問題,該文提出基于端到端的神經(jīng)網(wǎng)絡(luò)模型DF-RCNN以提高車輛密集區(qū)域的檢測精度。首先,在特征提取階段,DF-RCNN模型將深淺層特征圖的分辨率統(tǒng)一并融合;其次,DF-RCNN模型結(jié)合可變形卷積和可變形感興趣區(qū)池化模塊,通過加入少量的參數(shù)和計(jì)算量以學(xué)習(xí)目標(biāo)的幾何形變。實(shí)驗(yàn)結(jié)果表明,該文提出的模型針對密集區(qū)域的車輛目標(biāo)具有較好的檢測性能。
-
關(guān)鍵詞:
- 遙感影像 /
- 車輛檢測 /
- 密集區(qū)域 /
- 端到端卷積神經(jīng)網(wǎng)絡(luò)
Abstract: Vehicle detection is one of the hotspots in the field of remote sensing image analysis. The intelligent extraction and identification of vehicles are of great significance to traffic management and urban construction. In remote sensing field, the existing methods of vehicle detection based on Convolution Neural Network (CNN) are complicated and most of these methods have poor performance for dense areas. To solve above problems, an end-to-end neural network model named DF-RCNN is presented to solve the detecting difficulty in dense areas. Firstly, the model unifies the resolution of the deep and shallow feature maps and combines them. After that, the deformable convolution and RoI pooling are used to study the geometrical deformation of the target by adding a small number of parameters and calculations. Experimental results show that the proposed model has good detection performance for vehicle targets in dense areas. -
表 1 3 種尺度和 3 種比例映射候選區(qū)域尺寸
區(qū)域尺寸和比例 202, 1:1 202, 1:2 202, 2:1 302, 1:1 302, 1:2 302, 2:1 402, 1:1 402, 1:2 402, 2:1 映射區(qū)域大小 20×20 14×28 28×14 30×30 21×42 42×21 40×40 28×56 57×28 下載: 導(dǎo)出CSV
表 2 不同層檢測性能比較
模型 檢測區(qū)域 召回率 準(zhǔn)確率 F1指標(biāo) 層1 密集區(qū)域 0.461 0.611 0.525 層3 密集區(qū)域 0.623 0.754 0.682 層5 密集區(qū)域 0.655 0.923 0.766 層3+5 密集區(qū)域 0.714 0.930 0.808 層2+4 密集區(qū)域 0.709 0.925 0.803 層1+3+5 密集區(qū)域 0.725 0.927 0.814 下載: 導(dǎo)出CSV
表 3 結(jié)合可變形模塊檢測性能比較
模型 檢測區(qū)域 召回率 準(zhǔn)確率 F1指標(biāo) 層1+3+5,可變形卷積 密集區(qū)域 0.731 0.932 0.819 層1+3+5,可變形RoI池化 密集區(qū)域 0.726 0.925 0.814 層1+3+5,可變形卷積,可變形RoI池化 密集區(qū)域 0.744 0.940 0.831 層5,可變形卷積,可變形RoI池化 密集區(qū)域 0.725 0.924 0.812 層2+4,可變形卷積,可變形RoI池化 密集區(qū)域 0.715 0.922 0.805 層3+5,可變形卷積,可變形RoI池化 密集區(qū)域 0.729 0.928 0.817 下載: 導(dǎo)出CSV
表 4 不同模型檢測性能比較
模型 檢測區(qū)域 召回率 準(zhǔn)確率 F1指標(biāo) Faster-RCNN 密集區(qū)域 0.655 0.932 0.766 DCNN 密集區(qū)域 0.402 0.925 0.560 DF-RCNN 密集區(qū)域 0.744 0.940 0.831 Faster-RCNN 非密集區(qū)域 0.901 0.942 0.921 DCNN 非密集區(qū)域 0.962 0.785 0.865 DF-RCNN 非密集區(qū)域 0.910 0.952 0.931 下載: 導(dǎo)出CSV
-
SHAO Wen, YANG Wen, LIU Guang, et al. Car detection from high-resolution aerial imagery using multiple features[C]. Geoscience and Remote Sensing Symposium, Munich, Germany, 2012: 4379–4382. MORANDUZZO T and MELGANI F. Automatic car counting method for unmanned aerial vehicle images[J]. IEEE Transactions on Geoscience&Remote Sensing, 2013, 52(3): 1635–1647 doi: 10.1109/TGRS.2013.2253108 KLUCKNER S, PACHER G, GRABNER H, et al. A 3D teacher for car detection in aerial images[C]. International Conference on Computer Vision, Rio de Janeiro, Brazil, 2007: 1–8. TUERMER S, KURZ F, Reinartz P, et al. Airborne vehicle detection in dense urban areas using HoG features and disparity maps[J]. IEEE Journal of Selected Topics in Applied Earth Observations&Remote Sensing, 2013, 6(6): 2327–2337 doi: 10.1109/JSTARS.2013.2242846 CHEN Ziyi, WANG Cheng, WEN Chenglu, et al. Vehicle detection in high-resolution aerial images via sparse representation and superpixels[J]. IEEE Transactions on Geoscience and Remote Sensing, 2015, 54(1): 103–116 doi: 10.1109/TGRS.2015.2451002 康妙, 計(jì)科峰, 冷祥光, 等. 基于棧式自編碼器特征融合的SAR圖像車輛目標(biāo)識別[J]. 雷達(dá)學(xué)報(bào), 2017, 6(2): 167–176 doi: 10.12000/JR16112KANG Miao, JI Kefeng, LENG Xiangguang, et al. SAR target recognition with feature fusion based on stacked autoencoder[J]. Journal of Radars, 2017, 6(2): 167–176 doi: 10.12000/JR16112 王思雨, 高鑫, 孫皓, 等. 基于卷積神經(jīng)網(wǎng)絡(luò)的高分辨率SAR圖像飛機(jī)目標(biāo)檢測方法[J]. 雷達(dá)學(xué)報(bào), 2017, 6(2): 195–203 doi: 10.12000/JR17009WANG Siyu, GAO Xin, SUN Hao, et al. An aircraft detection method based on convolutional neural networks in high-resolution SAR images[J]. Journal of Radars, 2017, 6(2): 195–203 doi: 10.12000/JR17009 田壯壯, 占榮輝, 胡杰民, 等. 基于卷積神經(jīng)網(wǎng)絡(luò)的SAR圖像目標(biāo)識別研究[J]. 雷達(dá)學(xué)報(bào), 2016, 5(3): 320–325 doi: 10.12000/JR16037TIAN Zhuangzhuang, ZHAN Ronghui, HU Jiemin, et al. SAR ATR based on convolutional neural network[J]. Journal of Radars, 2016, 5(3): 320–325 doi: 10.12000/JR16037 YANG Zi and PUN-CHENG L S C. Vehicle detection in intelligent transportation systems and its applications under varying environments: A review[J]. Image&Vision Computing, 2017, 69: 143–154 doi: 10.1016/j.imavis.2017.09.008 CHEN Xueyun, XIANG Shiming, LIU Chenglin, et al. Vehicle detection in satellite images by hybrid deep convolutional neural networks[J]. IEEE Geoscience&Remote Sensing Letters, 2014, 11(10): 1797–1801 doi: 10.1109/LGRS.2014.2309695 LI Hao, FU Kun, YAN Menglong, et al. Vehicle detection in remote sensing images using denoizing-based convolutional neural networks[J]. Remote Sensing Letters, 2017, 8(3): 262–270 doi: 10.1080/2150704X.2016.1258127 REN Shaoqing, HE Kaiming, SUN Jian, et al. Faster R-CNN: Towards real-time object detection with region proposal networks[J]. IEEE Transactions on Pattern Analysis&Machine Intelligence, 2017, 39(6): 1137–1149 doi: 10.1109/TPAMI.2016.2577031 DENG Zhipeng, SUN Hao, ZHOU Shilin, et al. Toward fast and accurate vehicle detection in aerial images using coupled region-based convolutional neural networks[J]. IEEE Journal of Selected Topics in Applied Earth Observations&Remote Sensing, 2017, 10(8): 3652–3664 doi: 10.1109/JSTARS.2017.2694890 ELMIKATY M and STATHAKI T. Car detection in aerial images of dense urban areas[J]. IEEE Transactions on Aerospace&Electronic Systems, 2017, 54(1): 51–63 doi: 10.1109/TAES.2017.2732832 DAI Jifeng, QI Haozhi, XIONG Yunwen, et al. Deformable convolutional networks[OL]. http://arxiv.org/abs/1703.06211.2017. LIU Derong, LI Hongliang, and WANG Ding. Feature selection and feature learning for high-dimensional batch reinforcement learning: A survey[J]. International Journal of Automation and Computing, 2015, 12(3): 229–242 doi: 10.1007/s11633-015-0893-y YANG Bin, YAN Junjie, LEI Zhen, et al. Convolutional channel features[C]. IEEE International Conference on Computer Vision, Santiago, Chile, 2015: 82–90. ZHONG Bin, ZHANG Jun, WANG Pengfei, et al. Jointly feature learning and selection for robust tracking via a gating mechanism[J]. Plos One, 2016, 11(8): e0161808 doi: 10.1371/journal.pone.0161808 SIMONYAN K and ZISSERMAN A. Very deep convolutional networks for large-scale image recognition[OL]. http://arxiv.org/abs/1409.1556.2014. HE Kaiming, ZHANG Xiangyu, REN Shaoqing, et al. Deep residual learning for image recognition[C]. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, United States, 2016: 770–778. ZEILER M D and FERGUS R.Visualizing and Understanding Convolutional Networks[C]. European Conference on Computer Vision, Zurich, Switzerland, 2014: 818–833. -