RIASSUNTO
In this paper, we introduce a convolutional neural network based on the state-of-the-art detector, named YOLOv2, to classify and detect images of fish. Meanwhile, we have adopted the relevant customization techniques to optimize the architecture of model. The Nature Conservancy Fisheries Monitoring datasets, which consist of 5787 fish images belonging 7 classes, is used to dramatically scale the monitoring of fishing activities on a data science competition of Kaggle platform. This automatic classification and detection system has achieved 0.912 Mean Average Precision (MAP) and frame rate of 28.3 Frames Per Second (FPS), satisfying the real-time requirement for application.