ResNET-based underwater fish identification with Data Augmentation

Abstract

The underwater environment is intricate, and light stands as one of the foremost influences. When the model trained with fish pictures under ideal conditions is applied to real underwater recognition, the influence of light will challenge the recognition accuracy, and it is costly to collect new underwater fish datasets. In this paper, Data Augmentation is applied to enhance the model's accuracy for underwater fish recognition. This paper begins by introducing the Residual Network (ResNet) network and Data Augmentation techniques. It then presents the research idea including how to validate underwater effects on the model and the effectiveness of Data Augmentation to improve underwater fish identification accuracy, modeling, analyzes data on model performance above and underwater and finally draws conclusions of influences of Data Augmentation on underwater fish identification based on analysis. According to the experimental results, it is concluded that applying Data Augmentation to model training can improve model’s accuracy in underwater fish recognition. The significance of this method is that it not only saves the cost of a new datasets, but also makes the model better adapted to underwater fish recognition.