Fish feeding intensity assessment method using deep learning-based analysis of feeding splashes

•A method for assessing fish feeding intensity from thumbnail images of feeding splashes is proposed.•Adaptive tuning of receptive fields based on a multi-kernel selection mechanism.•Development of a semi-supervised object detector using CAM pseudo-labels.•Grey-level and grey–gradient co-occurrence matrices used to extract water splash classification features. Assessment of fish feeding intensity provides effective feedback on fish starvation, which is important for improving feed utilisation and reducing water pollution. The vigorous gathering, jumping, or chasing of fish during feeding produces splashes of different sizes. The size of the splash reflects the intensity of fish feeding. However, the diverse splashes, reflected light spots on the surface of the water, and ripples created by fish movement pose inevitable problems for direct quantification methods. These methods require extensive data labelling and complex pre-processing to eliminate disturbances. Hence, this study proposes a method for assessing fish feeding intensity using a workflow involving several deep learning-based techniques such as an adaptive receptive field image feature extraction network based on a multi-kernel selection structure that becomes the teacher module for a semi-supervised splash detection module that uses a class activation map to create pseudo-labels for splashes. Feature extraction is done on the detected feeding splashes using a grey-level co-occurrence matrix and a grey–gradient co-occurrence matrix, which are then used to assess fish feeding intensity using a regression multi-classification model. The experimental results show that the feeding intensity assessment method proposed in this study achieves a mean accuracy of 98.22%, a false detection rate of 1.27%, and a missed detection rate of 2.75%; these results provide practical implications for the design of effective feeding systems that can improve feed utilisation, reduce water pollution, and enhance the welfare of farmed fish.