自動魚種影像辨識方法之開發

Development of an Imaging Approach for Automatic Fish Identification

董真(Chen Tung)

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(2017年)

https://doi.org/10.6342/NTU201703469

摘要

為符合國際規範，提升我國的國際地位，確保漁民捕魚權益，我國遠洋漁業必須符合國際組織的相關規範，包括隨船配置觀察員，以蒐集紀錄漁獲與漁船作業相關資訊。然而人工觀測紀錄經常有不正確之現象，其精確性和公正性受到質疑。故本研究之目的在於發展一自動辨識魚獲種類之方法，做為日後發展電子觀察員之基礎。利用捲積類神經網路，並使用遷移學習之方法，希望建立用來分類台灣主要遠洋漁獲之分類器，其辨識對象分為鮪魚、旗魚、鯊魚和其他。研究過程中引用了兩個預先訓練之模型，並進行實驗測試捲積層固定數目、架構微調和參數優化。實驗結果顯示，以VGG16為基底之分類器便是正確率可達97.5%，而以Inception-V3為基底之分類器辨識率可達96.75%，兩者皆有良好的表現。

關鍵字

捲積類神經網路 ； 魚獲 ； 辨識 ； 影像分類 ； 遷移學習

並列摘要

In recent years, international organizations have regulated fishery in public seas to conserve marine ecosystems. Automatically identifying the types of fish catch is one of the most impartial approaches for monitoring fishing practices. Hence, this study proposes the use of deep learning algorithms to automatically identify major fish catch of Taiwan. Through the use of transfer learning, two convolutional neural network classifiers were developed to differentiate fish into four classes: tuna, billfish, shark, and others. During this process, two pre-trained models were employed as the base network and then fine-tuned to improve the identification ability when facing fine-grained-image classification problems. The experimental results show that these four classes of fish can be identified with a relatively high degree of accuracy (97.5% for the VGG16-based classifier and 96.75% for the Inception-V3-based classifier).

並列關鍵字

Convolutional neural networks ； fish catch ； identification ； image classification ； transfer learning

參考文獻

Adams, D. C., Rohlf, F. J., & Slice, D. E. (2004). Geometric morphometrics: ten years of progress following the ‘revolution’. Italian Journal of Zoology, 71(1), 5-16.

Alexandre, H., Vrignaud, J., Mangin, B., & Joly, S. (2015). Genetic architecture of pollination syndrome transition between hummingbird-specialist and generalist species in the genus Rhytidophyllum (Gesneriaceae). PeerJ, 3, e1028.

Cireşan, D. C., Meier, U., Gambardella, L. M., & Schmidhuber, J. (2010). Deep, big, simple neural nets for handwritten digit recognition. Neural computation, 22(12), 3207-3220.

Ge, Z., McCool, C., Sanderson, C., & Corke, P. (2015, September). Modelling local deep convolutional neural network features to improve fine-grained image classification. In Image Processing (ICIP), 2015 IEEE International Conference on (pp. 4112-4116). IEEE.

Hahnloser, R. H., Sarpeshkar, R., Mahowald, M. A., Douglas, R. J., & Seung, H. S. (2000). Digital selection and analogue amplification coexist in a cortex-inspired silicon circuit. Nature, 405(6789), 947-951.