Small object detection network for monitoring mass wasting activity in the Martian North polar region

Oleksii Martynchuk,Lida Fanara,Juergen Oberst

crossref（2022）

引用 0|浏览2

暂无评分

摘要

The north polar region of Mars is one of the most active places of the planet with avalanches[1] and ice block falls[2] being observed every year on High Resolution Imaging Science Experiment (HiRISE) data. Both phenomena originate at the steep icy scarps surrounding the north polar plateau, Planum Boreum. The exposed layers of ice and dust contain important information about the climate cycles of the planet[3]. We are interested in monitoring how the current scarp erosion rate (quantified through analyzing blockfalls) is linked to these periodic changes in temperature. The large scale of the region of interest, combined with a growing amount of available satellite data makes automation key for this project. Our goal is to robustly monitor the whole north polar region for ice block falls throughout the entire Mars ReconnaissanceOrbiter (MRO) mission.    Deep Learning networks are widely used for object detection problems, the YOLO family being particularly popular among researchers and commercial users. Although the baseline performance of its most recent iteration (v5) is satisfactory for most generic applications, HiRISE images of NPLD (North Polar Layered Deposits) present a greater challenge. Large boulders are usually easy to identify due to their well differentiated visual features, e.g. longer shadows and edges, greater surface area and less crowded backgrounds. As the blocks get smaller features  become less pronounced and margins between different object instances shrink beyond the default resolution of  YOLOv5 thus noticeably hindering the network’s overall efficiency. To boost the detection rate we propose the following modifications to the currently available build (fig1):   <ul> <li aria-level="1">Super Resolution module[4] - before the images are fed into the network backbone they are stretched out to 640x640p using simple bicubic interpolation. An alternative approach utilizes DASR (domain-distance aware super-resolution)  as a preprocessing step. This super-resolution model can be trained on analogous Earth data (e.g. Svalbard) and then deployed to intelligently upscale the HiRISE dataset, adding an extra layer of textures and sharpening the images. </li> </ul>   <ul> <li aria-level="1">Better anchor grid[5] - anchor grid parameters should be based on the size of the objects we expect to observe in our dataset. The closer the initial guess, the less adjustments are subsequently  performed by the network. We suggest that automatically deriving the anchors prior to training is extremely beneficial to the overall performance. </li> </ul>   <ul> <li aria-level="1">Additional Fusion Layer and Residual Connections[6] - baseline version of  YOLOv5 starts implementing residual connections only at P3 level, ignoring the higher resolution P2 features. To reduce the loss of the feature information of small objects we incorporate the P2 layer into the rest of the network by stacking it on top of the basis formed by the original three layers, i.e. {P3, P4, P5}. </li> </ul>   <ul> <li aria-level="1">Attention Module[6] - this enhancement helps the network learn the spatial correlation between features much like the human visual mechanism, which enables the eye to focus on more important feature regions and suppresses irrelevant regions. There are different options available, yet we use a combination of ECA-Net and CBAM contributing channel attention and spatial attention mechanisms respectively.</li> </ul>   <ul> <li aria-level="1">Position of attention modules within the network[6] - the insertion point of the aforementioned modules is not location invariant. We have chosen to incorporate it into the fusion layers right after every CSP (i.e.C3) block. </li> </ul> <img src="" alt="" /> Figure 1. A diagram of the improved YOLOv5 architechture. The input tesnors are already upscaled by DASR and presprocessed to derive the best posible anchor grid parameters. C3 (CSP) -  cross-stage partial network; SPPF - spatial pyramid pooling fast; ECBAM - ECANet + CBAM attention modules.  The boost in performance is most noticeable in ice boulder dense areas, where detection is highly dependent on resolving separate compactly stacked objects. The exact figure is yet to be confirmed experimentally but preliminary testing shows at least a 3% improvement. We hypothesize that integrating the DASR module with the main training routine instead of performing upscaling as a separate preprocessing step might boost the performance even further. The NPLD remains our primary area of interest due to its high levels of activity and good satellite image density, yet we also plan to apply our pipeline to different surface changes and Martian regions as well as other celestial objects. <ul> <li aria-level="1">Russell P., Thomas, N. Byrne, S. Herkenhoff, K.Fishbaugh, K. Bridges, N. Okubo, C. Milazzo, M. Daubar, I. Hansen. Seasonally active frost-dust avalanches on a north polar scarp of Mars captured by HiRISE, Geophys. Res. Lett. 35, L23204 (2008).</li> <li aria-level="1">L. Fanara, K. Gwinner, E. Hauber, J. Oberst. Present-day erosion rate of north polar scarps on Mars due to active mass wasting, Icarus, Volume 342, (2020),113434, ISSN 0019-1035.</li> <li aria-level="1">Smith I.B., Putzig N.E., Holt J.W. and Phillips R.J. An ice age recorded in the polar deposits of Mars, Science, 352, 1075–1078 (2016).</li> <li aria-level="1">Yunxuan Wei, Shuhang Gu, Yawei Li, Longcun Jin. Unsupervised Real-world Image Super Resolution via Domain-distance Aware Training (2021).</li> <li aria-level="1">Zhan, W., Sun, C., Wang, M. et al. An improved Yolov5 real-time detection method for small objects captured by UAV. Soft Comput 26, 361–373 (2022)</li> <li aria-level="1">Linlin Zhu, Xun Geng, Zheng Li and Chun Liu. Improving YOLO v5 with Attention Mechanism for Detecting Boulders from Planetary Images. Remote Sens. , 13(18), 3776 (2021).</li> </ul>

查看译文

AI 理解论文

溯源树

样例

生成溯源树，研究论文发展脉络

Chat Paper

正在生成论文摘要