Research on fish identification in tropical waters under unconstrained environment based on transfer learning

Abstract Seafloor unconstrained environment video is shot in uncontrolled open sea area. There are multiple backgrounds, complex illumination and weather changes, rapid growth of algae attached lenses, which affect the stability shooting quality, resulting difficulty image recognition. At present, there still no algorithm that superior to other methods generally, it necessary build a model combined with specific scenes applications. In this paper, fish identification method tropical waters based on transfer learning under proposed. Firstly, pre-processed by affine transformation realize data enhancement. Furthermore, RestNet50 deep convolutional neural network constructed compare effect recognition before after learning. The results show that, accuracy loss indicators better than those non-transfer when trained imagenet introduced as initial weight network. When 150 epochs, begin converge, can complete task environment.