利用深度學習量化白蝦進食相關特性

Abstract

蝦是全球主要的蛋白質來源。在蝦類養殖中，飼料成本大約占總支出的40%。有效的進食管理對於優化蝦的成長和最小化成本至關重要。蝦的食慾受到生長階段和環境條件的影響。此外，由於蝦類是底棲生物，使得直接觀察變得充滿挑戰。傳統上，蝦的食慾是通過將樣本飼料放置在傘網上進行肉眼觀察來確定的，但此方法耗時且主觀。為了解決這些問題，本研究旨在通過使用深度神經網絡觀察蝦類的餵食相關行為來自動化蝦的食慾判定。 在提議的方法中，構建了配備飼料投料器的水下攝影模組，以在樣本餵食過程中捕捉蝦隻的影片。通過影像處理算法對影片中的飼料殘留區域進行量化，並計算出反映蝦食慾的飼料殘留區域變化指數（FRAVI）。影片中的蝦隻被YOLOv9-c模型和追蹤演算法追蹤。接著，利用飼料殘留檢測模組進行測量，衍生出關鍵的進食相關特性，包括蝦類數量、移動、進入頻率和停留時間。飼料殘留檢測模組達到了0.885的整體相關性，而YOLOv9-c模型達到了0.88的平均精度。此外，還監測了水溫、鹽度和溶解氧等環境因素，分析它們與進食相關特性的相關性。分析表明，水溫與蝦類活動水平正相關，較高的站壓與蝦類進入頻率正相關，顯示這些因素在飼料攝取效率中扮演著重要角色。本研究提供了關於蝦隻進食相關行為的持續、客觀和精確的信息，這些資訊可能有助於農民優化飼料管理和水產養殖實踐。

Shrimp serves as a significant protein source globally. In shrimp farming, feed accounts for approximately 40% of the overall expenses. Effective feeding management is crucial for optimizing shrimp growth and minimizing the costs. Shrimp appetite is influenced by growth stages and ambient conditions. In addition, shrimps are benthos, making direct observation challenging. Conventionally, shrimp appetite was determined using naked-eye observation by putting sample feed on trays. The approach is, however, time-consuming and subjective. To address these issues, this study aimed to automate shrimp appetite by observing their feeding-related behaviors using deep neural networks. In the proposed approach, underwater video modules with feed dispensers were built to capture videos of shrimps during sample feeding (i.e., a small amount of feed). Feed residue areas in the videos were quantified using image processing algorithms. Feed residue area variation index (FRAVI) that indicates shrimp appetite was quantified. Shrimps in the videos were detected and tracked using YOLOv9-c and simple online realtime tracking algorithm. Feed residue were measured using feed residue detection module. Key feeding-related characteristics, including shrimp count, movement, entry frequency, dwelling time, were next derived. Feed residue detection module achieved an overall correlation of 0.885. YOLOv9-c model achieved a mean average precision of 0.88. Additionally, environmental factors like water temperature, salinity, and dissolved oxygen levels were monitored to analyze correlations with feeding-related behavior. Analysis indicated that water temperature is positively correlated with shrimp activity levels, and higher station pressure is positively correlated with shrimp entry frequency, suggesting these factors play a significant role in feed intake efficiency. The proposed approach provides continuous, objective, and precise information of the feeding-related behaviors of shrimps. The information may aid farmers in optimizing feed management and aquaculture practices.