Fish Image Classification Using Deep Convolutional Neural Network
Xiaojuan Lan, Juyang Bai, Meng Li, Jiajun Li
Abstract
Scientific research on species composition and geographical distribution of marine organisms is of great significance to the research of marine resources and the protection of rare species of marine life. In these studies, divers or underwater robots are often used to collect biological images, which are then manually classified by relevant experts. Manual-based classification is not only time-consuming but also prone to misjudge. Deep learning algorithms have also been applied in this field, but the classification performance is poor in general, mainly due to the low image quality and the small number of collected images. In response to this challenging, a fish classification algorithm based on Inception-V3 is proposed in this paper. First, data augmentation is realized by scaling, inverting, and panning of original images. Then transfer learning method is applied to improve the prediction accuracy. Experimental results show that the proposed method can effectively improve the classification accuracy, reaching about 89% for fish species.