目前人類的交流與生活使得海上船隻活動相當盛行，小至海港漁業大至國際貿易與物流，因此，在海巡監控上極需重視海上船隻位置與航線管理的重要性。在太空科技的蓬勃發展下，衛星可用來進行大規模地球探測活動，進而取代人力來提高效率以及降低成本。為了能有效地獲得衛星影像的探索目標資訊，我們使用深度學習演算法提取船隻特徵，以電腦視覺方法進行分割衛星影像中的船隻，相較於過去的船隻監控方法，我們所提的模型僅需分析衛星可見光影像這一種資料，且影像沒有地域及時間的限制，還能依照訓練的資料來擴充偵測的物件種類。
本研究提出一套透過衛星遙測影像進行海上船隻實例分割的方法；首先，對衛星影像進行強化與預判斷的前處理，以減少目標物尋找時間提升偵測準確性。接著，採取基於殘差卷積的網絡提取出衛星影像中的船隻特徵，並且結合具有注意力網路的特徵金字塔網路，進而增加深度學習船隻特徵的強韌性，以及有效去除背景雜訊的特徵；最後，以感興趣區域網路對高興趣區域進行衛星影像中的船隻實例分割。
在本論文以Airbus Ship Detection Challenge所提供的公開資料集進行研究與實驗，依據遮罩重合度和混淆矩陣的計算方法，其中以結合聚合殘差變換神經網路模型與注意力網路在進行衛星影像中的船隻分割，重合度0.5為基準，沒有預判斷分類模型下的像素分割平均精確度達87.96%，相較於實驗中的其他實例分割模型，我們的模型有最高的AP@50平均精確度；再者，於競賽上平均F2分數為0.8489，此外加上預判斷衛星影像存在船隻與否後再進行船隻偵測，其平均F2分數為0.8512。在僅使用單一模型與其他同類模型相比，我們的模型有最高的平均F2分數；而實驗結果顯示加上預判斷，此方法可有助於獲取更有效的衛星影像資訊。

At present, human communication and life make marine vessel activities quite frequent, ranging from harbor fishery to international trade and logistics. Therefore, it is necessary to pay attention to the marine vessel location and route management in maritime patrol monitoring. With the booming development of space technology, satellites can be used to conduct large-scale earth exploration activities, which can replace manpower to improve efficiency and reduce costs. In order to increase the effectiveness of the satellite and effectively obtain the exploration target information, we use deep learning algorithms to extract ship features and use computer vision methods to segment the ships in the satellite image. Compared with the past ship monitoring methods, the model we propose only needs to analyze the data of the satellite visible image, and the image has no geographical and time restrictions, and the types of detected objects can be expanded according to the training data.
This research proposes a method for instance segmentation of maritime ships through satellite images. First, the satellite imagery is augmented by data preprocessing and add the pre-determination step to reduce the target finding time and improve the detection accuracy. Second, the residual convolution network is used to extract the ship features in the satellite image. Then the Feature Pyramid Network combined with the attention network is used to increase the robustness of deep learning ship features and effectively remove the feature of background noise. Finally, the region-of-interest network is used to instance segment ships that are in the high-interest region of a satellite image.
This thesis conducts research and experiment on the public dataset provided by the Airbus Ship Detection Challenge and adopts the intersection over the union of mask and the confusion matrix as an evaluation standard. The combined aggregated residual transformations neural network model with the attention network is used to perform ship segmentation in satellite images. The pixel segmentation AP@50 of the model without the pre-determination step is 87.96%. Compared with other instance segmentation models in the experiment, our model achieves the highest AP@50; moreover, the average F2-score in the challenge is 0.8489. Also, we add the pre-determination to classify whether the satellite image contains the ship or not before detection, the average F2-score is 0.8512. Compared with other similar models using only a single model, our model has the highest average F2-score; and the experimental results show that with pre-determination, this method can obtain more effective satellite image information.

[1] L. Mou and X. X. Zhu, “Vehicle instance segmentation from aerial image and video using a multitask learning residual fully convolutional network,” IEEE Transactions on Geoscience and Remote Sensing, vol. 56, no. 11, pp. 6699-6711, 2018.
[2] Arkarjun, “Remote sensing illustration,” [Online]. Available: https://en.wikipedia.org/wiki/File:Remote_Sensing_Illustration.jpg. [Accessed May 1, 2021].
[3] S. Wei et al., “HRSID: A high-resolution SAR images dataset for ship detection and instance segmentation,” IEEE Access, vol. 8, pp. 120234-120254, 2020.
[4] S.Wei et al., “Precise and robust ship detection for high-resolution SAR imagery based on HR-SDNet,” Remote Sensing, vol. 12, no. 1, pp. 1-29, 2020.
[5] D. Hordiiuk et al., “Semantic segmentation for ships detection from satellite imagery,” in the Proceedings of the 39th IEEE International Conference on Electronics and Nanotechnology, pp. 454-457, 2019.
[6] O. Ronneberger, P. Fischer, and T. Brox, “U-net: convolutional networks for biomedical image segmentation,” in the Proceedings of the Medical Image Computing and Computer-Assisted Intervention, Springer, LNCS, vol. 9351, pp. 234-241, 2015.
[7] K. He et al., “Deep residual learning for image recognition,” in Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, Las Vegas, Nevada, pp. 770-778, 2016.
[8] Y. Chen, J. Zheng, and Z. Zhou, “Airbus ship detection-traditional v.s. convolutional neural network approach,” [Google Scholar]. Available: http://cs229.stanford.edu/proj2018/report/58.pdf. [Accessed May 1, 2021].
[9] X. Wu, D. Sahoo, and S. C. H. Hoi, “Recent advances in deep learning for object detection,” arXiv preprint arXiv:1908.03673, 2019.
[10] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet classification with deep convolutional neural networks,” in Proceedings of the Advances in Neural Information Processing Systems, Lake Tahoe, Nevada, pp. 1097-1105, 2012.
[11] M. Polat, H. Mahmood, and A. Mohammed, “Ship detection in satellite images,” in Proceedings of International Symposium on Applied Sciences and Engineering, Erzurum, Turkey, pp. 200-202, 2018.
[12] K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,” in Proceedings of International Conference on Learning Representations, San Diego, California, pp. 1-14, 2015.
[13] S. Ren et al., “Faster R-CNN: Towards real-time object detection with region proposal networks,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 39, no. 6, pp. 1137-1149, 2017.
[14] V. Badrinarayanan, A. Kendall, and R. Cipolla, “SegNet: A deep convolutional encoder-decoder architecture for image segmentation,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 39, no. 12, pp. 2481-2495, 2017.
[15] J. Long, E. Shelhamer, and T. Darrell, “Fully convolutional networks for semantic segmentation,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Boston, Massachusetts, pp. 3431-3440, 2015.
[16] Z. Huang et al., “Mask scoring R-CNN,” in Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, Long Beach, California, pp. 6402-6411, 2019.
[17] D. Bolya et al., “YOLACT: Real-time instance segmentation,” in Proceedings of the IEEE International Conference on Computer Vision, Seoul, Korea (South), pp. 9156-9165, 2019.
[18] A. Géron, “Introducing capsule networks,” [Online]. Available: https://www.oreilly.com/content/introducing-capsule-networks. [Accessed May 1, 2021].
[19] K. He et al., “Mask R-CNN,” in Proceedings of the IEEE International Conference on Computer Vision, Venice, Italy, pp. 2980-2988, 2017.
[20] M. Tan and Q. V. Le, “EfficientNet: Rethinking model scaling for convolutional neural networks,” in Proceedings of the 36th International Conference on Machine Learning, Long Beach, California, pp. 10691-10700, 2019.
[21] M. Sandler et al., “MobileNetV2: Inverted residuals and linear bottlenecks,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Salt Lake, Utah, pp. 4510-4520, 2018.
[22] J. Hu, L. Shen, and G. Sun, “Squeeze-and-excitation networks,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Salt Lake, Utah, pp. 7132-7141, 2018.
[23] S. Xie et al., “Aggregated residual transformations for deep neural networks,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, Hawaii, pp. 1492-1500, 2017.
[24] C. Szegedy et al., “Going deeper with convolutions,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Boston, Massachusetts, pp. 1-9, 2015.
[25] T. Y. Lin et al., “Feature pyramid networks for object detection,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, Hawaii, pp. 936-944, 2017.
[26] S. Woo et al., “CBAM: Convolutional block attention module,” in Proceedings of the Conference on European Conference on Computer Vision, Munich, Germany, pp. 3-19, 2018.
[27] Airbus, “Airbus Ship Detection Challenge | Kaggle,” [Online]. Available: https://www.kaggle.com/c/airbus-ship-detection/overview. [Accessed May 1, 2021].
[28] S. Karki and S. Kulkarni, “Ship detection and segmentation using Unet,” in Proceedings of the International Conference on Advances in Electrical, Computing, Communication and Sustainable Technologies, Bhilai, India, pp. 1-7, 2021
[29] Z. Huang, S. Sun, and R. Li, “Fast single-shot ship instance segmentation based on polar template mask in remote sensing images,” in Proceedings of International Geoscience and Remote Sensing Symposium, Virtual Symposium, pp. 1236-1239, 2020.
[30] X. Nie et al., “Attention Mask R-CNN for ship detection and segmentation from remote sensing images,” IEEE Access, vol. 8, pp. 9325-9334, 2020.
[31] Airbus, “Airbus Ship Detection Challenge | discussion 74443 | Kaggle,” [Online]. Available: https://www.kaggle.com/c/airbus-ship-detection/discussion/74443. [Accessed May 1, 2021].
[32] Airbus, “Airbus Ship Detection Challenge | discussion 71782 | Kaggle,” [Online]. Available: https://www.kaggle.com/c/airbus-ship-detection/discussion/71782. [Accessed May 1, 2021].
[33] Airbus, “Airbus Ship Detection Challenge | discussion 71667 | Kaggle,” [Online]. Available: https://www.kaggle.com/c/airbus-ship-detection/discussion/71667. [Accessed May 1, 2021].