利用深度卷積類神經網路偵測及計算影片中魚體並測量魚體長

Detecting and Counting Harvested Fish and Measuring Fish Body Lengths in EMS Videos Using Deep Convolutional Neural Networks

曾啟軒(Chi-Hsuan Tseng)

指導教授 ： 郭彥甫

國立臺灣大學/生物資源暨農學院/生物產業機電工程學研究所/碩士(2019年)

https://doi.org/10.6342/NTU201903433

摘要

捕撈漁獲的統計是海洋資源永續利用及管理的關鍵因素，近年來有許多船隻已經利用電子觀察員系統(EMS)來記錄漁船作業情況，接著觀察員在資料中心判讀EMS的影片並做出捕撈魚穫的統計，人工的判讀和記錄既費時又消耗大量人工，因此，本研究提出利用深度卷積神經網路自動偵測及計算影片中魚體並測量魚體長的方法，在研究中以遮罩區域卷積神經網路(Mask R-CNN)偵測並切割影片每幀中的魚體，利用時間和距離閥值計算魚體數量，接著以Mask R-CNN預測的機率和遮罩來辨識魚的類別和測量魚體長，本研究的Mask R-CNN模型在魚體偵測上達到96.46%的召回率及93.51%的平均精確率，本研究的魚體計算方法達到93.84%的召回率及77.31%的精確率，本研究在影片中魚類別辨識達到98.06%的準確率。

關鍵字

卷積類神經網路 ； 魚體長 ； 漁業資源管理 ； 物件偵測 ； 實體切割

並列摘要

The statistics of harvested fish are key indicators for marine resource management and sustainability. In recent years, electronic monitoring systems (EMS) are used to record the fishing practices of vessels. The statistics of the harvested fish in the EMS videos later are manually read and collected by the operators in data centers. Manual collection is, however, time consuming, and labor intensive. This study proposes to automatically detect harvested fish, identify fish types, and measure fish body lengths in the EMS videos using deep learning. In the study, the fish in the frames of the EMS videos were detected and segmented from the background at pixel level using mask regional-based convolutional neural networks (Mask R-CNN). The counting of the fish was then determined using time thresholding and distance thresholding. Subsequently, the types and body lengths of the fish were next determined using the confidence scores and the masks, respectively, predicted by the Mask R-CNN model. The developed Mask R-CNN model reached a recall of 96.46% and a mean average precision of 93.51% in detection. The proposed method for fish counting reached a recall of 93.84% and a precision of 77.31%. The proposed method for fish type identification reached an accuracy of 98.06%.

並列關鍵字

Convolutional neural networks ； fish body length ； fish resource management ； object detection ； instance segmentation

參考文獻

Abadi, M., Barham, P., Chen, J., Chen, Z., Davis, A., Dean, J., ... & Kudlur, M. (2016). Tensorflow: A system for large-scale machine learning. In 12th {USENIX} Symposium on Operating Systems Design and Implementation ({OSDI} 16) (pp. 265-283).

Ames, R. T., Leaman, B. M., & Ames, K. L. (2007). Evaluation of video technology for monitoring of multispecies longline catches. North American Journal of Fisheries Management, 27(3), 955-964

Chollet, F. (2015). Keras. https://github.com/fchollet/keras

Everingham, M., & Winn, J. (2011). The PASCAL visual object classes challenge 2012 (VOC2012) development kit. Pattern Analysis, Statistical Modelling and Computational Learning, Tech. Rep.

FAO. (2017). Seafood traceability for fisheries compliance: Country-level support for catch documentation schemes. Rome: FAO.