Abstract: Submarine video system is the most intuitive investigation method for marine scientific research, especially for investigating the types of seafloor rocks and the bottom environment. By interpreting high-quality video and picture data obtained from deep-sea high-quality cameras and photographic equipments, it is possible to directly identify the types of seafloor rocks, biological activities and seabed environments. In recent years, with the widespread use of deep-sea camera equipments such as manned submersibles, Remotely Operated Vehicles (ROVs), and deep-sea optical tow vehicles, scientists has acquired more and more submarine high-definition videos and image data. However, large workload is needed to process these videos and image data. Therefore, it is necessary to develop some intelligent seabed image recognition technologies based on computer vision. In this study, based on the deep learning technology, an automatic seabed image recognition model is proposed, and the model is applied to the processing of video data obtained from the Mid-Atlantic Ridge hydrothermal sulfide survey. Firstly, based on high-quality video data obtained from oceanographic survey, 31 499 frames of seabed images were manually identified and labeled as pelagic sediments, pillow basalts, breccia basalts, hydrothermal sulfides and other categories, and hydrothermal sulfide is the main exploration target. This image dataset is randomly segmented into either the training set or the validation set of the deep learning model. Secondly, deep residual networks (ResNet) were built, and image datasets were used to be training and verifying accuracy. Finally, this model was used to analyze a 3.5 km long submarine camera survey line, and the results show that the ResNet model identification accuracy rate reaches 98%. This method has the comprehensive advantages of high efficiency and high precision for the intelligent recognition of submarine video data, and can be used not only for post-processing of massive video data, but for on-site analysis of deep-sea video surveys.