Research of Surface Floating Object Detection Method Based on Improved YOLOv4

This paper proposes an improved YOLOv4 water surface floating object detection method to address the problem that floating objects on the water surface are difficult to detect due to their small size and existing detection algorithms are computationally intensive and difficult to run in real time on embedded devices. The specific idea is to use the MobileNetV3 network to replace the backbone network in YOLOv4; at the same time, negative feedback is introduced, and the proportion of the loss of small-sized targets in the overall loss during the training process is used as feedback, so that selective data enhancement can be performed to improve the detection accuracy of small-sized targets. The experimental results show that compared with the traditional YOLOv4 algorithm, the model parameter number of the proposed algorithm is reduced by 82.4% and the detection speed is improved by 52%.