Fish Behavior Tracking Algorithm Based on Multi-Domain Deep Convolutional Neural Network

Abstract

Fish behavior can be used as an important indicator of the water quality in mariculture. Therefore, changes in fish behavior can reflect changes in water quality in a timely and effective manner. In order to track fish behavior, a fish target location and behavior tracking algorithm based on multi-domain deep convolutional neural network is proposed. Firstly, the cross-entropy loss function is used to distinguish the target and background in each domain of the multi-domain deep convolutional neural network, and the loss function is optimized by the stochastic gradient descent method (SGD) to find the local minimum of the loss function. Then the hard example mining method is used to resolve the positive and negative sample imbalance problem, and the loss function convergence is accelerated. Finally, the Bounding-Box regression algorithm is used to adjust the position of the fish tracked by the model, which accurately locates the fish target. The experimental results demonstrate that our proposed scheme achieve better tracking accuracy compared with the CNN-SVM scheme.