Fish feeding intensity quantification using machine vision and a lightweight 3D ResNet-GloRe network

Abstract

Quantifying feeding intensity of fish is important in developing intelligent feeding control system, thus improving feed utilization rate and reducing water pollution. The current study explored a real-time, high-precision and lightweight 3D ResNet-Glore fish feeding intensity quantification network, which can accurately locate the four levels of fish feeding intensities in video stream. In this network, the lightweight GloRe module is expanded in 3D space, and the Residual block in the 3D ResNet network is modified to form the 3D GloRe module. The relational reasoning is achieved through graph convolution in the interactive space to improve accuracy of discrimination. In addition, the sliding window and the frame extraction processing of the video data significantly reduces the model parameters and the amount of calculation. Experimental results showed that the classification accuracy for four types of feeding intensity was 92.68%, which is 4.88% higher compared with that of the classical 3D ResNet network. The parameters were decreased by 46.08% and the GFLOPs decreased by 44.10%. The proposed network improved the training and recognition speed and reduced the hardware equipment requirements, which can provide a theoretical basis for subsequent feeding decisions.