基于深度卷积神经网络的水稻纹枯病检测识别

Detection and Recognition of Rice Sheath Blight Based on Deep Convolutional Neural Network

纹枯病是水稻的三大病害之一,尤其在中国北方稻区,纹枯病发生逐渐加重、严重威胁到中国的粮食安全,而纹枯病的有效检测是水稻病害预防与控制的首要任务。在实际生产中,农民和从事相关的研究人员通过人工目测来识别水稻纹枯病,但由于光线、杂草、枯叶等外在自然因素和人眼视觉误差等人为因素,导致对水稻的病害等级误判,从而影响对水稻纹枯病的防治,造成环境污染和经济损失,而计算机视觉技术给水稻纹枯病的自动识别检测带来了可能。基于2019年沈阳农业大学北方粳型超级稻成果转化基地的水稻纹枯病图像数据,综合借鉴YOLOv1、YOLOv2和Faster R-CNN算法,设计了一种基于深度卷积神经网络的水稻纹枯病识别模型:YRSNET。该模型具有回归思想的特点,将图像划分为相同大小互不重合的网格,然后通过特征图来预测每个网格区域上的边界框和含有纹枯病病斑的置信度,最终通过非极大值抑制法获得含有纹枯病病斑的最佳边界框位置。试验结果表明:YRSNET对纹枯病病斑识别的平均精度mAP为84.97%、查准率达到为90.21%,对大小为450×800pixel的图像识别所需时间为32.26ms(31帧·s^-1),可满足复杂背景下的水稻植株图像纹枯病的检测,对智能农业水稻纹枯病有效防治具有重要意义。

Sheath blight is one of the three major diseases of rice,especially in northern China,where the occurrence of sheath blight is gradually increasing,which seriously threatens China’s food security.Effective detection of sheath blight is the primary task of rice disease prevention and control.In actual production,farmers and related researchers use manual visual inspection to identify rice sheath blight.However,due to external natural factors such as light,weeds,dead leaves and human visual errors,the gradem misjudgment of prevention and control for rice sheath blight resulted in,environmental pollution and economic losses.Computer vision technology makes it possible to automatically identify and detect rice sheath blight.Based on the rice sheath blight image data of the Northern Japonica Super Rice Achievement Transformation Base of Shenyang Agricultural University in 2019,this research comprehensively draws on the YOLOv1,YOLOv2 and Faster R-CNN algorithms to design a rice sheath blight based on deep convolutional neural network Disease recognition model:YRSNET.The model has the characteristics of regression thinking.It divides the image into grids of the same size that do not overlap with each other,and then uses feature maps to predict the bounding box and the confidence of the sheath blight spots on each grid area.The maximum value suppression method obtains the best bounding box position containing the sheath blight spots.The test results showed that the average accuracy mAP of YRSNET for spot identification of sheath blight disease was 84.97%,the corresponding precision was 90.21%,and the time required to recognize an image with a size of 450×800 pixel is 32.26 ms(31 frames·s^-1),which can meet the detection of rice sheath blight in the image of rice plants in a complex background.So,it is of great significance for the effective control of rice sheath blight in smart agriculture.