Abstract:
Deep learning models perform well in open-air captured image classification problems but, in the case of underwater images, can’t find better accuracy due to low-level features. Under-water image classification is a difficult task due to the independent nature of the sea. To overcome low-level issues, we need to implement advanced deep learning models. Meanwhile, the deep learning community focuses on pre-trained deep networks to classify out-of-domain images and transfer learning. This study proposes a model for underwater image classification. We split the Labeled Fishes in the Wild (LFW) dataset into two versions as raw and enhanced before implementation. Strategy II of the transfer learning approach has been used by analyzing the nature of the dataset. We have applied pre-trained CNN architectures such as VGG, ResNet, Xception, and DenseNet on the dataset to find better results. DenseNet121 offers the most accurate predictions for raw and enhanced versions. We achieved 97.84% classification accuracy on the original version and gained 99.35% accuracy on the enhanced version dataset. In the end, we tested our application on Fish Species Image Dataset (FSID) and achieved correct results
