基于卷积神经网络的葡萄叶片病害检测方法

摘要

文章采用多角度建议区域Faster-RCNN准确定位图像中葡萄叶片,提出一种基于卷积神经网络的病害检测方法,检测图像叶片病害.相比直接检测图像病害,可去除背景因素对病害区域干扰,降低错误率.结果表明,该算法对自然条件下葡萄病害成像适应性良好.文章统计6种不同条件下拍摄图像,对一般叶片检测算法平均mAP为75.52％,显著高于传统算法.在病害检测时,采用两种策略:从一幅图像中检测到每个单个叶片,或将整幅图像对叶片取掩模后,作为下一级病害检测器输入图像.结果表明,第一种方法,6种常见葡萄病害平均mAP为66.47％,其中褐斑病与白粉病mAP超过70％;第二种方法,病害检测平均mAP为51.44％,但平均检测时间节约75％.两种方法性能均优于在原始图像上直接病害检测方法.%A method of disease detection based on Convolutional neural network was proposed in this paper.Firstly,the grape leaves were located by faster-RCNN using the proposals from multiple angles,then,detected the diseases over regions of the leaves.Compared with detecting diseases in the image,this method improved the accuracy of disease detection observably.The experimental results showed that the proposed algorithm achieved good performance for handling the grape images taken under natural conditions.For leaf detection,our method got average mAP 75.52％ and surpassed other three algorithms as the comparison.To detect disease,two different strategies had been introduced.The first was that each leaf in an image was cut out and fed to the disease detector,respectively.The second was that original image was masked on the detected leaves and taken as the input for the detector.Statistic results showed that the average mAP of the first method achieves 66.47％ for six common diseases.Among them,for the first method,mAPs of brown spot and powdery mildew were over 70％.For the second method,the average mAP was 51.44％.However,it saved about 75％ time cost than the first one.Furthermore,the performances of the two methods were superior to the method detecting diseases on an image directly.