Improved YOLOv4 for Water Wastes Detection
Gu Shaokui
School of Mechanical Engineering, North China University of Water Resources and Electric Power, Zhengzhou, Henan 450000, China and Henan Key Laboratory of Underwater Intelligent Equipment (713th Research Institute of China State Shipbuilding Corpration Limited), Zhengzhou 40015, China.
Niu Jinxing *
School of Mechanical Engineering, North China University of Water Resources and Electric Power, Zhengzhou, Henan 450000, China.
Li Longyan
School of Mechanical Engineering, North China University of Water Resources and Electric Power, Zhengzhou, Henan 450000, China.
*Author to whom correspondence should be addressed.
Abstract
Quantifying plastic refuse in water area helps to understand how plastic refuse accumulates in water area and is essential for targeted cleanup efforts. Currently, the most common methods for quantifying plastic in water area are human visual counting and sampling using nets, but such methods are costly and labor-intensive. This study proposes a watershed refuse identification algorithm based on an improved YOLOv4. Lightweight improvements to YOLOv4. EfficientNetB1 is used to replace the backbone network of YOLOv4, and the Depthwise Convolution is used to replace the original convolution to reduce the number of model parameters and computation. The anchors are re-clustered using k-means algorithm to improve the accuracy. The experimental results show that the improved algorithm improves the detection speed by 11.2% and reduces the number of parameters by 76.54% compared with YOLOv4 at the expense of 0.69% recognition accuracy.
Keywords: Deep learning, refuse detection, YOLOv4, EfficientNet