Abstract Fine-grained visual classification is one of the essential data science tasks with enormous datasets. The studies show that species composition and ample distribution of fishes notably impact the fishery industry, aquaculture, and marine ecosystem. Incredible work and analysis are required to state fish characteristics by classification. Lately, deep learning has helped to gain exceptional development in this area. Be that as it may, fine-grained fish classification is more complex than primary image classification, particularly with medium quality (i.e., underwater images) and small-scale (i.e., limited data). But traditional convolutional neural networks (CNNs) and other popular models like V.G.G., RESNET, DenseNet, etc., require high-quality and high-scale datasets. This paper presents another way to enhance the CNN models that best fit this fine-grained fish classification problem. Real-world underwater images have several issues, including noise, dominant colours, light attenuation, etc. Further, it isn't easy to get a large set of images of each category of species under the sea, and hence an imbalanced dataset is generated. These two problems are addressed in this paper. Then the quality of the raw images was improved by an Underwater Image Enhanced Generative Adversarial Network (UIEGAN), that CycleGAN trains over 6128 images of the ImageNet dataset. Conventional data augmentation helps increase the dataset size of the dataset by random transformations of the images (i.e., flipping, rotation), but it cannot handle the imbalanced class problem. We generated synthetic images of every class utilizing DCGAN to create a balanced dataset. Further, we used the SmallerVGG and SmallerRESNET models that best fit the Croatian dataset. Moreover, we compared our strategy with eight popular pre-trained transfer learning models trained on the ImageNet dataset. The exploratory outcomes show that the proposed techniques beat well-known CNNs, with high accuracy, demonstrating their possible applications in the real-time underwater fish image classification.
Field : Mühendislik
Journal Type : Uluslararası
Relevant Articles | Author | # |
---|
Article | Author | # |
---|