Knowing precise fish appetite is a prerequisite for developing a high-efficient feeding system in aquaculture. However, the current studies on the assessment of fish appetite mostly focus on relevant spatial features of fish school, ignoring the time series-based variation characteristics in the process of fish feeding, which may decrease the accuracy of appetite assessment. To address the research gap and solve these problems, a novel and efficient fish appetite grading method, based on the spatial-temporal characteristics of fish behavior, was proposed in this study, using the modified kinetic energy model and customized recurrent neural network. First, the modified kinetic energy model was used to quantify and extract the behavioral spatial characteristics of fish school without foreground segmentation and individual tracking. The temporal features of fish feeding behavior were learned based on the vector sequence of spatial characteristics above, by means of a customized recurrent neural network. Following this, fish appetite level was determined with the help of layers of full connection and softmax. Through the exhaustive test on four different behavior datasets, the presented method shows better performance (accuracy: 97.08%, 97.35%, 92.50%, 98.31%, respectively) on appetite assessment of fish than many other state-of-the-art methods.

了解精确的鱼类食欲是在水产养殖中开发高效投喂系统的先决条件。然而，目前对鱼类食欲评价的研究多集中在鱼群的相关空间特征上，忽视了鱼类摄食过程中基于时间序列的变化特征，这可能会降低食欲评价的准确性。为了弥补研究空白​​并解决这些问题，本研究基于鱼类行为的时空特征，使用改进的动能模型和定制的递归神经网络，提出了一种新颖高效的鱼类食欲分级方法。首先，在没有前景分割和个体跟踪的情况下，使用改进的动能模型量化和提取鱼群的行为空间特征。基于上述空间特征向量序列，通过定制的循环神经网络学习鱼类摄食行为的时间特征。在此之后，在全连接层和 softmax 层的帮助下确定鱼的食欲水平。通过对四个不同行为数据集的详尽测试，所提出的方法在鱼类食欲评估方面表现出比许多其他最先进的方法更好的性能（准确度：分别为 97.08%、97.35%、92.50%、98.31%）。