Fish recognition from a vessel camera using deep convolutional neural network and data augmentation

Nowadays, as a sub topic of computer vision and fishery industry, fish recognition is still a challenging work not only because of biodiversity, but also because of the complex background of images. In this paper, we aim to classify different fish images obtained from cameras of fishing vessels. Fish classification from a vessel camera is different with the best known and the most well-investigated object classification problem. Fish always take only a small part of the whole image. For these two reasons, it's a challenge for us to detect and classify fish froma vessel. Our work is done with Kaggle dataset. The Kaggle dataset aims to detect and classify the species of fish. Eight categories are available in the dataset. In order to overcome these problems we introduced two methods to improve our accuracy for the fish classification. We propose a local region based the fish area modeling approach to force the model focus on the fish region. We also apply an elastic rotation-based data augmentation method to avoid over-fitting which may be caused by the imbalanced training dataset. We employ the AlexNet, GoogLeNet, Caffenet and VGGNet neural network for the classification. The experimental results show our method successfully enhanced the classification performance.