Find in Library

Sea cucumber culture and fishing are primarily dependent on manual work. For fast and accurate automatic identification of sea cucumbers, deep residual networks with different configures were conducted in this experiment to identify underwater sea cucumber. Sea cucumber images were captured by a C-Watch remotely operated underwater vehicle (ROV) in a sea cucumber fishery at Haiyang Qiandao Lake in Shandong Province, China and sliced to positive samples and negative samples. Two training algorithms, namely, the stochastic gradient descent algorithm (SGD) and Adam, activation functions ReLU and leaky ReLU, as well as learning rates of 0.001, 0.005, 0.01, 0.05, and 0.1 were combined to form different models, which were trained with epochs 200 times and mini-batch of 100. The results showed that the accuracy of each model was higher than 82%, and the highest accuracy reached 89.53% under the SGD algorithm with ReLU and a learning rate of 0.05 or 0.1, which showed better generalization ability than that of other models. The performance of the proposed method indicates a great potential for automatic sea cucumber identification. Keywords: Sea cucumber, Identification, Convolutional neural networks, Deep residual networks