International Journal of Scientific Research and Engineering Development

Coral reefs are vital ecosystems that support marine biodiversity but are under severe threat. Accurately identifying corals using underwater images is crucial for monitoring their health; however, this task is challenging due to issues such as color distortion caused by water depth and turbidity. Existing methods, primarily based on Convolutional Neural Networks (CNNs), often struggle with these color variations and fail to capture the broader ecological context of a reef. This paper presents a Dual-Branch CNN-Transformer Hybrid Model that synergistically combines a CNN and a Vision Transformer (ViT) designed to classify coral and non-coral images in underwater environments. This model combines the local feature-detection strength of a CNN (ResNet18) with the global context-understanding ability of a Pre-trained Vision Transformer (ViT). Synthetic color-shift transformations have been applied to simulate various underwater conditions, such as light scattering and water turbidity.
The model was evaluated on a dataset of 1690 images consisting of 1300 coral images and 390 noncoral images. and achieved an overall test accuracy of 93.49%. The results demonstrate that our hybrid approach is a highly effective and reliable tool for automated coral identification, highlighting the potential of deep learning and computer vision in marine conservation and biodiversity protection.