As it is hard to recognize marine fishes occurred in submarine observation videos due to the bad undersea environment and low quality of the video, a recognition method based on deep learning was proposed. Firstly, the video was split into pictures, and as this type of video contains a large proportion of useless data, a background subtraction algorithm was used to filter the pictures without fish to save the time of processing all data. Then, considering the undersea environment is blurring with low bright, based on the dark channel prior algorithm, the pictures were preprocessed to improve their quality before recognition. Finally, a recognition deep learning model based on Convolutional Neural Network (CNN) was consructed with weighted convolution process to improve the robustness of the model. The experimental results show that, facing submarine observation video frames with poor quality, compared with traditional CNN, the method with preprocessing and weighted convolution as hidden layer can increase the recognition accuracy by 23%, contributing to the recognition of marine fishes in submarine observation video.