Research on Small Object Detection in Complex Marine Environments

With the improvement of computer arithmetic power, deep learning is evolving rapidly. Utilizing deep learning-based algorithms to the detection of floating objects at marine and marine rescue can not only save human and material resources but also improve efficiency. However, given the complex marine environments, it is hard for existing algorithms to achieve satisfactory results. In this paper, the SRB-YOLO detection method is proposed to add an image preprocessing module during the training process and input the images before and after preprocessing into the backbone network for training, to enhance the stability of the model, and improve the feature fusion network of the basic YOLOv5, which improves the detection result and increases the accuracy by seven percentage points.