基于改进YOLOv4的林业害虫检测

为了提高林业害虫检测的准确性,提出一种基于YOLOv4的改进算法。首先,基于智能害虫捕捉装置拍摄的图像,制作害虫数据集,采用K-means算法对样本数据集的目标框进行聚类分析,基于DIoU-NMS算法实现对害虫的计数功能;然后,在模型的路径聚合网络（PANet）结构上增加特征融合和104×104层级特征检测图,以提升对小个体害虫的识别率;最后,根据模型检测效率和复杂度,调整模型中的尺度特征图组合,在保证检测准确度的基础上,提升检测效率,并精简模型。试验结果表明,改进的YOLOv4模型的平均识别精度比传统YOLOv4模型提高了1.6百分点,且对于小个体害虫的识别效果更好,模型复杂度和模型参数量分别减少了11.9%、33.2%,检测速度提升了11.1%,更适于应用部署。

Detection of forest pests based on improved YOLOv4

Abstract：In order to improve the accuracy of pest detection in the forest, an improved algorithm based on YOLOv4 was proposed. First, based on the image captured by the intelligent pest capture device, the pest data set was made, and K-means algorithm was used to cluster the target frame of the sample data set. Based on the DIoU-NMS algorithm, the counting function of pests was realized. Then, feature fusion was added to the path aggregation network (PANet) structure of the model, as well as 104×104 hierarchical feature detection graph, to increase the recognition accuracy of small-sized pests. Finally, according to the efficiency and complexity of model detection, the combination of scale feature maps in the model was adjusted to ensure the detection accuracy, improve the detection efficiency and simplify the model. The experimental results showed that the mean average precision of the improved YOLOv4 model was 1.6 percent higher than that of the traditional YOLOv4 model with a better performance on the detection of small-sized pests. Besides, its speed was improved by 11.1 percent, and the model complexity and model parameters were reduced by 11.9 percent and 33.2 percent, respectively, as compared with the traditional YOLOv4 model, which was more suitable for application deployment.