Coral morphology detection in underwater imagery using YOLOv12 with CNN and transformer encoder fusion

For monitoring marine biodiversity and observing the health of reefs, it's important to be able to accurately see coral morphology in underwater images. Present approaches for finding underwater objects still have problems because light can be absorbed, scattered, or become turbid, and coral morphologies look very similar to each other, which often results in misclassification, erratic localization, and lower accuracy. The present automated detectors, like YOLOv7 and latest newer variants, like YOLOv11, do not work well with underwater image distortions, do not reason well about global features, or are not efficient enough to be used in practical marine monitoring systems. This paper introduces an innovative method for detecting coral morphology automatically from underwater images, utilizing YOLOv12 augmented with a Convolutional Neural Network (CNN) and transformer encoder integration in real-time to mitigate these constraints. The suggested model uses the local feature extraction power of CNN and the global context modelling power of transformers to improve feature representation in intricate underwater situations. The model is more accurate and efficient because it uses the strength of YOLOv12 along with the strengths of CNN and transformer encoders to detect objects. Empirical assessments on a standard underwater dataset show that our method is better than the present models in terms of precision, recall, and mean Average Precision (mAP), while still being able to make inferences in real time. This work helps make coral reef monitoring tools better by giving them a reliable and scalable way to identify objects underwater.