Fig 1.
Experimental system diagram.
Fig 2.
RGB images of three different feeding states.
(A) None Feeding State. (B) Weak Feeding State. (C) Strong Feeding State.
Table 1.
Classification criteria of feeding state intensity.
Fig 3.
Model structure diagram.
Fig 4.
Optical flow images of three different feeding states.
(A) None Feeding State. (B) Weak Feeding State. (C) Strong Feeding State.
Fig 5.
The basic structure of ResNet-50.
Fig 6.
ResNet network residual unit structure.
Fig 7.
Binary images of three different feeding states.
(A) None Feeding State. (B) Weak Feeding State. (C) Strong Feeding State.
Fig 8.
Basic structure of the one-dimensional convolutional neural network.
Fig 9.
The training and testing results for the temporal feature network (A), spatial feature network (B), and data statistical feature network (C).
Table 2.
Network training results.
Fig 10.
Three confusion matrix diagrams depicting the test results based on temporal feature network (A), spatial feature network (B), and data statistical feature network (C).
Table 3.
Evaluation index of classification results of fish feeding status by three-stream network.
Table 4.
Ablation study.
Table 5.
Comparison with other models for fish feeding behavior discrimination.
Fig 11.
Class activation mapping (CAM) of test samples.
(A), (B), and (C) are samples with correct predictions, classed as None, Weak, and Strong. (D) denotes the misclassified sample. Its essential actual value is Weak, and the predicted values are None.
Table 6.
Comparison of processing time of different feature network models.
Fig 12.
Software function interface.