Abstract: For the demand of autonomous recognition for small objects in subsea operations by remote operated vehicles (ROVs), a YOLOv3 algorithm based on deep learning is proposed. ROVs are trained according to neural network weight of the data collected in ROV fetching competition, in order to identify and classify small underwater objects fast and precisely, which will solve the underwater-object-recognition problem in complex and unknown subsea operation environment. The effect of algorithm learning rate on loss functions in the seafood data set is analyzed. The results show that underwater object recognition and classification of seafood based on YOLOv3 algorithm is of high real time as well as precision. The precision ratio of all objects classified reaches up to ninetynine percent, and the recall ratio is above ninety percent. The estimating speed can reach up to 35 frames per second(FPS). It also shows that adjusting learning rate during network training process can help decrease loss functions.