Robust Underwater Fish Classification Based on Data Augmentation by Adding Noises in Random Local Regions

Underwater fish classification is in great demand, but the unrestricted natural environment makes it a challenging task. The monitor placed underwater gets a lot of low-quality and hard-to-mark marine fish images. These images suffer from various illumination, complex background etc. At the same time, there are many high-quality and easy-to-mark marine fish pictures on the Internet. In this paper, we propose an effective data augmentation approach for improving the classification accuracy of low-quality marine fish images. In our method, unlike the existing global image method, random local regions are proposed for simulating local occlusion and fuzziness in various underwater environment. In addition, four types of noise are incorporated for augmenting training data set. Experimental results demonstrate that our approach can significantly enhance the classification performance of low-quality marine fish images under various challenging conditions when using high-quality marine fish images as training sets.