| With the continuous development of society,more and more domestic garbage enters the water bodies and floats on the water surface under the influence of man-made or natural factors.As one of the factors leading to the pollution of water bodies,these garbage floating materials seriously affect the ecological environment of water bodies as well as threaten the drinking water safety of residents.Real-time monitoring of water pollution is an important part of the environmental management of waters and floating prevention and control,the traditional reliance on manual inspection of the regulatory approach is inefficient and difficult to achieve real-time monitoring.The relatively complex water surface environment increases the difficulty of floating object target detection,in which the detection of small targets is prone to miss detection,false detection,etc.For this reason,this paper proposes to improve the water surface floating object detection method of YOLOv5(You only look once),using common floating objects such as bottles,plastic bags,water plants and dead fish as detection objects,and constructing a water surface floating object target detection model.By realizing the high-precision real-time detection of floating objects on the water surface,it is of practical significance to build a smart water conservancy system and promote the construction of smart water conservancy.The specific research work of this paper is as follows:1.To address the problems of insufficient data and low quality of floating objects on the water surface,we collected relevant images and manually labeled each floating object target in the images by various means to produce a larger and higher quality floating object data set in a fixed format.2.To improve the detection of small targets floating objects,a data enhancement method for small targets is proposed to improve the detection capability of the network for small targets by increasing the number of small targets in the data as well as increasing the complexity of the training background.3.To further improve the detection performance of the YOLOv5 object detection algorithm,the K-means++ algorithm is used to cluster the prior frames and improve the matching of the prior frames with the data features.An efficient coordinate attention mechanism is introduced in the feature extraction network to enhance the focus of the network on the target.The neck feature fusion network is reconstructed based on the Bi FPN idea using a simple residual structure to further enhance the fusion of features at different levels.Using SIo U＿Loss as the localization loss function,the model can converge faster and improve the prediction box regression accuracy by introducing angular loss.Through performance comparison experiments in the surface floating object test dataset,the results showed that the improved algorithm improved the average detection accuracy by 9.7% and decreased the error detection rate by 8.5%,among which the small target error detection rate decreased by 7.3%.4.The embedded device based on Rexchip RV1126 chip is used as the model deployment platform,and the relevant operators are optimized for the embedded device,while the network structure is adjusted during the model conversion.Finally,through model porting and deployment testing,high accuracy real-time detection of floating object targets on the water surface is achieved without partial frame loss and detection frame shift during the testing process. |
