Improved YOLOv5 s and transfer learning for floater detection.

This study aims to address the detection and classification of floating objects on water surfaces, including items such as bottles, plastic bags, aquatic plants, and dead fish, which pose significant threats to water quality and ecosystems. Traditional detection methods rely on manual observation and cleanup, which are inefficient, costly, and risky. To tackle this challenge, this paper proposes a solution based on an improved YOLOv5 s model by collecting floating object image data and constructing and processing the dataset using manual photography and SAGAN data augmentation techniques. We optimized the YOLOv5 s model by integrating the EfficientNetv2 lightweight network, the content-aware reassembly of features lightweight upsampling module, the bidirectional feature pyramid network structure, and by introducing attention modules such as squeeze-and-excitation and efficient multi-scale attention, along with the scylla intersection over union (SIoU) loss function. Additionally, transfer learning techniques were employed to enhance the model's performance in detecting floating objects on water surfaces, and ablation experiments were conducted to validate the effectiveness of each improvement. The results show that the improved YOLOv5 s model exhibits better performance and generalization ability on the test set, with a 5.27 percentage point increase in model accuracy. The model's parameter count, computational load, and weight size are 53.9%, 21.3%, and 54% of the original YOLOv5 s model, respectively, providing an efficient, accurate, and real-time solution for detecting floating objects on water surfaces. The methodology presented in this paper holds significant importance for the monitoring of aquatic ecological environments and the management of floating debris, offering valuable insights for achieving precise and efficient detection and classification of floating objects on water surfaces.