Chu-Chan LeeYan-Fu KuoYuan-Nan Chu2024-10-282024-10-282024https://www.scopus.com/record/display.uri?eid=2-s2.0-85206107928&origin=resultslisthttps://scholars.lib.ntu.edu.tw/handle/123456789/722510Shrimp is a major source of protein worldwide. In shrimp farming, the cost of feed accounts for approximately 40% of the overall expenses. Precise feeding management is the key to optimize the growth rate of shrimp while maintaining farming costs. Shrimp appetite is affected by numerous factors, including environmental conditions, growth stages, etc. Because the aforementioned factors are usually confounded, direct observation is the best approach to determine shrimp appetite. Conventionally, a small amount of feed (also referred to as trial feed) was put on trays. The trays were put into shrimp ponds for a period of time. Farmers then determined the amounts to feed shrimp by observing the characteristics (e.g., number of shrimp and amount of trial feed residue) on the trays. Manual observation was, however, discontinuous and time-consuming. The interpretation of the characteristics of shrimp could also vary between farmers. To address these issues, this study proposed to automatically and continuously quantify the characteristics of shrimp during trial feeding using underwater video systems (UVSs)and machine vision. UVSs were installed at the bottom of a pond, acquiring underwater videos during scheduled trial feeding events. Shrimp in the video was detected using You Only Look Once—Version 9 Compact (YOLOv9c) and tracked using simple online and real-time tracking. Characteristics of shrimp, including staying duration, number of shrimp, flowrate of visits, and average movement, that are strongly related to their appetite were then quantified. The trained YOLOv9c model achieved a mean average precision of 0.92. The proposed approach is fully automatic and objective. The characteristics quantified by the proposed approach may help farmers to optimize the feed management as well as shrimp farming aquaculture practices.falsecomputer visionConvolutional neural networkobject detectionobject trackingshrimpshrimp behavior[SDGs]SDG2[SDGs]SDG14conference paper10.13031/aim.2024010342-s2.0-85206107928