Fish School Feeding Behavior Quantification Using Acoustic Signal and Improved Swin Transformer

Abstract

In aquaculture, real-time quantification of fish feeding behavior is one of the important bases for feeding decisions. During feeding, the acoustic s produced by fish chewing feed and activities can be used to quantify the feeding behavior. Therefore, this paper proposes an ASST ( Audio Spectrum Swin Transformer ) model based on acoustic signal and attention mechanism, which can divide the feeding intensity of fish into four grades: strong, medium, weak, and none. The specific implementation methods are as follows: (1) A sliding window is applied to clip the audio, and the acoustic. signals are transformed into spectrograms. (2) The perceptual domain of fish feeding acoustic spectrum recognition task is gradually expanded by adopting the Swin Transformer and utilizing its hierarchical structure. (3) The model's performance for small data sets is improved by adding SPT, LSA, as well as enhanced residual connections. (4) A predictive optimization module is designed to correct the feeding strategy according to 4 feeding levels. The final experimental results show that the accuracy of the improved ASST network for fish feeding behavior quantification reaches 96.16%, and it can effectively identify four classes of fish feeding intensities, which can achieve on-demand feeding and provides the basis for developing intelligent feeding machines.

Keywords: Aquaculture, Feeding acoustics recognition, Fish feeding behavior quantification, Improved Swin Transformer network, Deep learning.