Benchmark dataset on feeding intensity of the pearl gentian grouper(Epinephelus fuscoguttatus♀×E. lanceolatus♂)

In aquaculture, the assessment of fish feeding intensity is an essential indicator of fish appetite, which is significant for guiding feeding and optimizing fish production. In the factory-based recirculating water high-density aquaculture environment, images have disadvantages such as uneven contrast and blurring, and there are difficulties in manually extracting image features. Although the deep learning-based fish feeding intensity assessment model has higher recognition accuracy and better robustness, the conventional differentiation of feeding intensity usually relies on manual experience to divide the feeding intensity dataset, which is subjective and uncertain, and the annotation is observed by the aquaculture experienced personnel to increase the labor and time cost. In order to solve these problems, this study constructs a benchmark dataset of the feeding intensity of pearl gentian groupers in a factory circulating water environment, which is divided into feeding fish groups and fish aggregation areas by training Unet semantic segmentation network and compares standard clustering algorithms through clustering evaluation indexes to maximally select the optimal clustering method and the number of clusters that are suitable for this paper's dataset. In addition, training on this benchmark dataset proposes an improved feeding intensity evaluation network, which achieves a good balance in prediction accuracy and parameter memory and offers the possibility of subsequent deployment of the model on mobile. Dataset address: https://github.com/Qhj-123/fig-dataset • Real-world scenario data: We collected images of fish feeding in high-density recirculating aquaculture environments. • Benchmark dataset: We constructed a benchmark dataset for the assessment of pearl gentian grouper feeding intensity. • Deployment network: We proposed an improved lightweight network based on MobileNetV3.