# American Institute of Mathematical Sciences

November  2019, 2(4): 299-314. doi: 10.3934/mfc.2019019

## Big Map R-CNN for object detection in large-scale remote sensing images

 a. FIST LAB, School of Information Science and Engineering, Yunnan University Kunming, 650091, Yunnan, China b. Yunnan Union Vision Technology Co Ltd. Kunming, 650091, Yunnan, China c. School of Software, Yunnan University Kunming, Yunnan University Kunming, 650091, Yunnan, China

* Corresponding author: Dapeng Tao

Published  December 2019

Detecting sparse and multi-sized objects in very high resolution (VHR) remote sensing images remains a significant challenge in satellite imagery applications and analytics. Difficulties include broad geographical scene distributions and high pixel counts in each image: a large-scale satellite image contains tens to hundreds of millions of pixels and dozens of complex backgrounds. Furthermore, the scale of the same category object can vary widely (e.g., ships can measure from several to thousands of pixels). To address these issues, here we propose the Big Map R-CNN method to improve object detection in VHR satellite imagery. Big Map R-CNN introduces mean shift clustering for quadric detecting based on the existing Mask R-CNN architecture. Big Map R-CNN considers four main aspects: 1) big map cropping to generate small size sub-images; 2) detecting these sub-images using the typical Mask R-CNN network; 3) screening out fragmented low-confidence targets and collecting uncertain image regions by clustering; 4) quadric detecting to generate prediction boxes. We also introduce a new large-scale and VHR remote sensing imagery dataset containing two categories (RSI LS-VHR-2) for detection performance verification. Comprehensive evaluations on RSI LS-VHR-2 dataset demonstrate the effectiveness of the proposed Big Map R-CNN algorithm for object detection in large-scale remote sensing images.

Citation: Linfei Wang, Dapeng Tao, Ruonan Wang, Ruxin Wang, Hao Li. Big Map R-CNN for object detection in large-scale remote sensing images. Mathematical Foundations of Computing, 2019, 2 (4) : 299-314. doi: 10.3934/mfc.2019019
##### References:

show all references

##### References:
Motivation for the proposed method. (a) Remote sensing scene of Madrid Airport. (b) Remote sensing scene of the South China Sea. These examples are from the RSI LS-VHR-2 dataset. The targets in the images are indicated by red cicles. The remote sensing scenes show the characteristics of large scale, high resolution, and relatively sparse target distribution, which means that existing methods are suboptimal for detection
The scheme of Big Map R-CNN, containing three main components: 1) cropping the input big map in the form of a sliding window; 2) detecting each sub-image sequentially and filtering possible object areas; 3) using mean shift clustering to precisely locate candidate object areas, cropping the new sub-images containing possible objects, and using quadric-detecting to judge whether there is an object or not
Large-scale image cropping
PRCs of the proposed Big Map R-CNN method and three other state-of-the-art detection methods (YOLOv3, Faster R-CNN, and Mask R-CNN). (a) is the PRC of the four methods for aircraft when IoU = 0.5; (b) is the PRC of the four methods for aircraft when IoU = 0.75; (c) is the PRC of the four methods for ships when IoU = 0.5; (d) is the PRC of the four methods for ships when IoU = 0.75
Some examples from the RSI LS-VHR-2 dataset
Detection comparisons of the different methods. (a) Typical Mask R-CNN for aircraft; (b) Big Map R-CNN for aircraft; (c) typical Mask R-CNN for ships; (d) Big Map R-CNN for ships. The true positives are indicated by green rectangles, the false negatives are indicated by red circles, and the bounding boxes that deviate from the ground truth are indicated by red rectangles
DESCRIPTION OF THE RSI LS-VHR-2 DATASET
 Label Name Total instances Complete instances Fragmentary instances Scene class Images Image width Sub-images 1 aircraft 103917 85975 17942 203 2858 6000-15000 62129 2 ship 68436 54386 14050 30 397 5000-18000 53860
 Label Name Total instances Complete instances Fragmentary instances Scene class Images Image width Sub-images 1 aircraft 103917 85975 17942 203 2858 6000-15000 62129 2 ship 68436 54386 14050 30 397 5000-18000 53860
DETAILS OF THE TEST IMAGES
 Label Scale(pixels) Images Instances Sub-images aircraft $8000\times8000$ 5 272 980 ship $8000\times8000$ 5 225 980
 Label Scale(pixels) Images Instances Sub-images aircraft $8000\times8000$ 5 272 980 ship $8000\times8000$ 5 225 980
PARAMETER SETTING OF Mask R-CNN AND Big Map R-CNN
 Input Size Per Batch Size Max Iteration Anchor Stride Base Learning Rate Steps Weight Decay NMS Threshold Momentum 600 8 90000 (4, 8, 16, 32, 64) 0.01 (60000, 80000) 0.0001 0.7 0.9
 Input Size Per Batch Size Max Iteration Anchor Stride Base Learning Rate Steps Weight Decay NMS Threshold Momentum 600 8 90000 (4, 8, 16, 32, 64) 0.01 (60000, 80000) 0.0001 0.7 0.9
PERFARMANCE COMPARISONS OF THREE DIFFERENT CROPPING SIZE IN Faster R-CNN NETWORK
 Cropping Size AP Cost time(s) C300 0.430 45.82 C600 0.651 13.20 C800 0.647 8.79
 Cropping Size AP Cost time(s) C300 0.430 45.82 C600 0.651 13.20 C800 0.647 8.79
PERFORMANCE COMPARISONS OF THE FOUR METHODS ON AIRCRAFT
 Method IoU=0.5 IoU=0.75 TP FP FN Recall Precision AP TP FP FN Recall Precision AP YOLOv3 213 25 59 0.783 0.895 0.727 166 72 106 0.610 0.6974 0.494 Faster R-CNN 242 55 30 0.890 0.815 0.830 189 108 83 0.695 0.636 0.618 Mask R-CNN 245 38 27 0.901 0.866 0.843 184 99 88 0.676 0.650 0.570 Big Map R-CNN 261 4 11 0.960 0.985 0.959 241 24 31 0.886 0.909 0.850
 Method IoU=0.5 IoU=0.75 TP FP FN Recall Precision AP TP FP FN Recall Precision AP YOLOv3 213 25 59 0.783 0.895 0.727 166 72 106 0.610 0.6974 0.494 Faster R-CNN 242 55 30 0.890 0.815 0.830 189 108 83 0.695 0.636 0.618 Mask R-CNN 245 38 27 0.901 0.866 0.843 184 99 88 0.676 0.650 0.570 Big Map R-CNN 261 4 11 0.960 0.985 0.959 241 24 31 0.886 0.909 0.850
PERFORMANCE COMPARISONS OF THE FOUR METHODS ON SHIP
 Method IoU=0.5 IoU=0.75 TP FP FN Recall Precision AP TP FP FN Recall Precision AP YOLOv3 128 53 97 0.569 0.707 0.513 66 115 159 0.293 0.365 0.213 Faster R-CNN 164 185 61 0.729 0.470 0.651 78 271 147 0.347 0.223 0.259 Mask R-CNN 166 121 59 0.738 0.578 0.661 78 209 147 0.347 0.272 0.273 Big Map R-CNN 191 49 34 0.849 0.796 0.826 133 107 92 0.591 0.554 0.546
 Method IoU=0.5 IoU=0.75 TP FP FN Recall Precision AP TP FP FN Recall Precision AP YOLOv3 128 53 97 0.569 0.707 0.513 66 115 159 0.293 0.365 0.213 Faster R-CNN 164 185 61 0.729 0.470 0.651 78 271 147 0.347 0.223 0.259 Mask R-CNN 166 121 59 0.738 0.578 0.661 78 209 147 0.347 0.272 0.273 Big Map R-CNN 191 49 34 0.849 0.796 0.826 133 107 92 0.591 0.554 0.546
THE AVERAGE PRECISION OF Mask R-CNN AND Big Map R-CNN IN RSI LS-VHR-2 DATASET
 Method Backbone AP($\%$) Mask R-CNN ResNet50 75.2 Big Map R-CNN ResNet50 89.2
 Method Backbone AP($\%$) Mask R-CNN ResNet50 75.2 Big Map R-CNN ResNet50 89.2
COMPREHENSIVE PERFORMANCE COMPARISONS OF FOUR METHODS
 Method mAP (IoU=0.5) mAP (IoU=0.75) Inference time(s/im) YOLOv3 0.620 0.354 3.310 Faster R-CNN 0.741 0.439 13.254 Mask R-CNN 0.752 0.422 13.310 Big Map R-CNN 0.892 0.700 16.005
 Method mAP (IoU=0.5) mAP (IoU=0.75) Inference time(s/im) YOLOv3 0.620 0.354 3.310 Faster R-CNN 0.741 0.439 13.254 Mask R-CNN 0.752 0.422 13.310 Big Map R-CNN 0.892 0.700 16.005
 [1] Zhihua Zhang, Naoki Saito. PHLST with adaptive tiling and its application to antarctic remote sensing image approximation. Inverse Problems & Imaging, 2014, 8 (1) : 321-337. doi: 10.3934/ipi.2014.8.321 [2] Boris Kramer, John R. Singler. A POD projection method for large-scale algebraic Riccati equations. Numerical Algebra, Control & Optimization, 2016, 6 (4) : 413-435. doi: 10.3934/naco.2016018 [3] Danuta Gaweł, Krzysztof Fujarewicz. On the sensitivity of feature ranked lists for large-scale biological data. Mathematical Biosciences & Engineering, 2013, 10 (3) : 667-690. doi: 10.3934/mbe.2013.10.667 [4] Mahmut Çalik, Marcel Oliver. Weak solutions for generalized large-scale semigeostrophic equations. Communications on Pure & Applied Analysis, 2013, 12 (2) : 939-955. doi: 10.3934/cpaa.2013.12.939 [5] Philippe Bonneton, Nicolas Bruneau, Bruno Castelle, Fabien Marche. Large-scale vorticity generation due to dissipating waves in the surf zone. Discrete & Continuous Dynamical Systems - B, 2010, 13 (4) : 729-738. doi: 10.3934/dcdsb.2010.13.729 [6] Fan Jia, Xue-Cheng Tai, Jun Liu. Nonlocal regularized CNN for image segmentation. Inverse Problems & Imaging, 2020, 14 (5) : 891-911. doi: 10.3934/ipi.2020041 [7] Rongliang Chen, Jizu Huang, Xiao-Chuan Cai. A parallel domain decomposition algorithm for large scale image denoising. Inverse Problems & Imaging, 2019, 13 (6) : 1259-1282. doi: 10.3934/ipi.2019055 [8] Tsuguhito Hirai, Hiroyuki Masuyama, Shoji Kasahara, Yutaka Takahashi. Performance analysis of large-scale parallel-distributed processing with backup tasks for cloud computing. Journal of Industrial & Management Optimization, 2014, 10 (1) : 113-129. doi: 10.3934/jimo.2014.10.113 [9] Suli Zou, Zhongjing Ma, Xiangdong Liu. Auction games for coordination of large-scale elastic loads in deregulated electricity markets. Journal of Industrial & Management Optimization, 2016, 12 (3) : 833-850. doi: 10.3934/jimo.2016.12.833 [10] Bo You, Chengkui Zhong, Fang Li. Pullback attractors for three dimensional non-autonomous planetary geostrophic viscous equations of large-scale ocean circulation. Discrete & Continuous Dynamical Systems - B, 2014, 19 (4) : 1213-1226. doi: 10.3934/dcdsb.2014.19.1213 [11] Boling Guo, Guoli Zhou. Finite dimensionality of global attractor for the solutions to 3D viscous primitive equations of large-scale moist atmosphere. Discrete & Continuous Dynamical Systems - B, 2018, 23 (10) : 4305-4327. doi: 10.3934/dcdsb.2018160 [12] Masataka Kato, Hiroyuki Masuyama, Shoji Kasahara, Yutaka Takahashi. Effect of energy-saving server scheduling on power consumption for large-scale data centers. Journal of Industrial & Management Optimization, 2016, 12 (2) : 667-685. doi: 10.3934/jimo.2016.12.667 [13] Bo You, Chunxiang Zhao. Approximation of stationary statistical properties of the three dimensional autonomous planetary geostrophic equations of large-scale ocean circulation. Discrete & Continuous Dynamical Systems - B, 2020, 25 (8) : 3183-3198. doi: 10.3934/dcdsb.2020057 [14] Rouhollah Tavakoli, Hongchao Zhang. A nonmonotone spectral projected gradient method for large-scale topology optimization problems. Numerical Algebra, Control & Optimization, 2012, 2 (2) : 395-412. doi: 10.3934/naco.2012.2.395 [15] Gaohang Yu. A derivative-free method for solving large-scale nonlinear systems of equations. Journal of Industrial & Management Optimization, 2010, 6 (1) : 149-160. doi: 10.3934/jimo.2010.6.149 [16] Bo You. Well-posedness for the three dimensional stochastic planetary geostrophic equations of large-scale ocean circulation. Discrete & Continuous Dynamical Systems, 2021, 41 (4) : 1579-1604. doi: 10.3934/dcds.2020332 [17] Yigui Ou, Wenjie Xu. A unified derivative-free projection method model for large-scale nonlinear equations with convex constraints. Journal of Industrial & Management Optimization, 2021  doi: 10.3934/jimo.2021125 [18] Hong Seng Sim, Chuei Yee Chen, Wah June Leong, Jiao Li. Nonmonotone spectral gradient method based on memoryless symmetric rank-one update for large-scale unconstrained optimization. Journal of Industrial & Management Optimization, 2021  doi: 10.3934/jimo.2021143 [19] Bo You. Optimal distributed control of the three dimensional primitive equations of large-scale ocean and atmosphere dynamics. Evolution Equations & Control Theory, 2021, 10 (4) : 937-963. doi: 10.3934/eect.2020097 [20] David Burguet. Examples of $\mathcal{C}^r$ interval map with large symbolic extension entropy. Discrete & Continuous Dynamical Systems, 2010, 26 (3) : 873-899. doi: 10.3934/dcds.2010.26.873

Impact Factor: