基于改进YOLOv5的农业害虫检测

Detection of Agricultural Pests Based on Improved YOLOv5 Algorithm

设计了一种改进的YOLOv5模型实现对小目标害虫识别定位并将其应用于农业虫情分析。首先对数据集中的图片进行数据增强操作,结合目标背景分析,对一半数据进行降低亮度处理以增大训练难度。为了解决目标检测模型中小目标容易随着卷积神经网络的加深而导致目标信息丢失的问题,提出在模型中额外增加特征提取层,以达到丰富小目标的特征信息和位置信息的目的。该模型采用一种切割图像再检测的方法,在降低模型计算量的同时提高了检测精度。最后设计了多组实验并进行对比,验证了模型的有效性,平均检测精度达92%。

An improved YOLOv5 model is designed to identify and locate small target pests for agricultural pest analysis.For the existing data set,firstly,the data enhancement operation was carried out on the images,and then the target background analysis was combined to reduce the brightness of half of the data to increase the training difficulty.To solve the problem that small and medium targets in the target detection model are easy to lose target information with the deepening of the convolutional neural network,a feature extraction layer was added to the model to enrich the feature information and location information of small targets in the target detection model.The model adopts a cutting image re-detection method,which reduces the calculation of the model and improves the detection accuracy.Finally,several experiments were designed and compared to verify the validity of the model,with an average detection accuracy of 92%.