Fish species recognition using VGG16 deep convolutional neural network

Conservation and protection of fish species is very important in aquaculture and marine biology. A few studies have introduced the concept of fish recognition; however, it resulted in poor rates of error recognition and conservation of a small number of species. This study presents a fish recognition method based on deep convolutional neural networks such as VGG16, which was pre-trained on ImageNet via transfer learning method. The fish dataset in this study consists of 50 species, each covered by 15 images including 10 images for training purpose and 5 images for testing. In this study, we trained our model on four different types of dataset: RGB color space image, canny filter image, blending image, and blending image mixed with RGB image. The results showed that blending image mixed with RGB image trained model exhibited the best genuine acceptance rate (GAR) value of 96.4%, following by the RGB color space image trained model with a GAR value of 92.4%, the canny filter image trained model with a GAR value of 80.4%, and the blending image trained model showed the least GAR value of 75.6%. © 2019. The Korean Institute of Information Scientists and Engineers.