Coral identification and quantification are vital for marine ecosystem conservation and coastal protection. To enable large-scale, semi-automatic coral analysis using images captured by Autonomous Underwater Vehicles (AUVs), we propose a hybrid deep learning framework that combines YOLOv9 for detection and Detectron2 for segmentation. The lightweight GhostNetv2 feature extraction network enhances YOLOv9’s backbone, while Dynamic Snake Convolution and Weighted Boxes Fusion (WBF) improve accuracy by refining feature extraction and eliminating overlapping low-confidence boxes. For semantic segmentation, Detectron2 leverages a ResNet50-based encoder-decoder architecture to extract coral features. Additionally, the Natural Underwater Image Color Enhancement (NUCE) technique mitigates underwater image distortion by recovering high-frequency details. Integrated into a user-friendly GUI, the system provides seamless coral detection and segmentation. Experiments show the enhanced YOLOv9 achieves an mAP@0.50 of 97%, and Detectron2 achieves 91%, demonstrating superior performance in lightweight, real-time scenarios. This approach offers a powerful tool for marine experts, advancing scalable coral monitoring and conservation.