YOLO-feed: An advanced lightweight network enabling real-time, high-precision detection of feed pellets on CPU devices and its applications in quantifying individual fish feed intake

Abstract

With the rapid advancement of deep learning, its application in aquaculture has grown significantly. However, the widespread use of deep learning models in this field is hindered by the large number of parameters and high computational power requirements. Moreover, accurately quantifying individual feed intake is essential for genetic improvement of feed efficiency, scientific feeding practices, and increasing aquaculture yield. Unfortunately, existing methods struggle to operate in real-time on low-performance hardware. To address these challenges, this study proposes YOLO-Feed, a cost-effective and highly efficient deep learning network. YOLO-Feed significantly reduces model size by eliminating the large object detection head and its corresponding branch network from You Only Look Once (YOLO) v8s, while introducing a powerful intersection-over-union (IoU) loss to enhance performance. We trained the model on over 9000 labeled feed pellet images. The YOLO-Feed model achieved a precision of 0.966, comparable to YOLOv8s, while reducing the number of parameters to less than 1 million— a 12.3-fold compression. In comparative experiments, YOLO-Feed maintained near-optimal precision with minimal parameters and was the only model capable of processing a single image on a CPU in just 16.9 ms. Furthermore, in an indoor aquaculture experiment involving 1027 fish, YOLO-Feed was successfully applied to measure individual feed intake, achieving an average accuracy of 90.3 % over a 733-min experimental period. These results demonstrate the superior performance and scalability of the YOLO-Feed model. To expedite research on the genetic improvement of feeding efficiency, we have open - sourced this project on GitHub (https://github.com/miaomiaoge/YOLO-Feed).