Computer Vision-Based Fish Feed Detection and Quantification System

Abstract

The development of the Automatic Feeder instrument and OAK-D camera has yielded positive results. The Automatic Feeder functions well, dispensing 30 grams of fish feed every 5 rotations of the stepper motor. The OAK-D camera records with sharp details, accurate colors, and good contrast, producing high-quality videos. The YOLOv5x detection model achieves an accuracy of 82%, precision of 80%, recall of 84%, mAP of 81.90%, and a training loss of 0.079144. This model can detect fish feed with high accuracy. The calculation of fish feed reveals different consumption patterns in the morning, afternoon, and evening. On average, the fish feed is depleted at the 25th minute across all time periods. The information from the graphs and tables can assist in optimizing the feeding process to avoid overfeeding.