Aquatic Animal Image Classification Technology Based on Transfer Learning and Data Augmentation

Yuan, H.; Zhang, S.; Qin, E., and Zhou, H., 2020. Aquatic animal image classification technology based on transfer learning and data augmentation. In: Hu, C. and Cai, M. (eds.), Geo-informatics and Oceanography. Journal of Coastal Research, Special Issue No. 105, pp. 129–133. Coconut Creek (Florida), ISSN 0749-0208. A method based on data augmentation and transfer learning methods is proposed to address the problems of complicated steps, low accuracy, and poor generalization of traditional aquatic animal image classification methods. First, a deep convolutional generative adversarial network is used to enhance the data of the sample. Second, based on modifying the source model's fully connected classification layer, the weights of high-level convolution modules are set to be trained for adaptive adjustment. Finally, using the accuracy on the validation set and training time as the evaluation indexes, the performance experiments were conducted on various network structures and a different proportion of trainable parameters. The experimental results show that when the portion of trainable parameters is around 75%, the accuracy of the image recognition model obtained by retraining can reach 97.9%, which can be improved by up to 36 percentage points compared with the support vector machine algorithm.