基于Python的大豆病害图像检测技术研究

Research on Soybean Disease Image Detection Technology Based on Python

为通过有效训练获得检测大豆病害的Python图像识别模型,保证样本数量、多样性和图片采集精度,利用Python爬虫技术编写大豆病害图像获取程序,结合数据扩充方法,在爬虫获取的目标图像的基础上扩充数据量,编写Python语言特征匹配程序对图片进行精确筛选。结果表明:利用爬虫技术对大豆病害图片进行采集可加快大豆花叶病、灰斑病、菌核病、霜霉病、根腐病、细菌性角斑病腐、枯萎病、炭疽病图像获取速度,提高数据集的多样性;经过局部二值模式处理后,结合纹理特征的差异,有效缩小了判断的范围,减小相似度判别的难度;通过均值哈希算法计算相似度后筛选的菌核病及枯萎病图像获取准确率均为100%,根腐病及灰斑病图像获取准确率最差,为83.3%,其他图像获取准确率均在90%以上,因此利用均值哈希算法计算相似度后,病害图片的获取准确性大大提高;通过Python语言编写的数据扩充代码经过图片旋转、翻转模糊、增加噪声、改变亮度几种处理,达到17倍扩增,试验最终获得2592张大豆病害图像。研究提高了大豆病害图像数据采集的精确度,为大豆常见病害的自动识别和诊断提供了技术参考。

In order to obtain the Python image recognition model for detecting soybean diseases through effective training,and ensure the sample size,diversity and image acquisition accuracy,the Python crawler technology is used to write the soybean disease image acquisition program,combined with the data augmentation method,expand the data amount on the basis of the target image obtained by the crawler,and write the Python language feature matching program to accurately screen the pictures.The results showed that the collection of soybean disease images by reptilian technology could accelerate the acquisition speed of soybean mosaic disease,gray spot,sclerotia,downy mildew,root rot,bacterial keratosis,wilt and anthrax,and improve the diversity of the dataset.After the local binary mode processing,combined with the difference of texture features,the scope of judgment was effectively reduced,and the difficulty of similarity discrimination was reduced.After calculating the similarity by the mean hash algorithm,the accuracy of the screened images of sclerotium and wilt was 100%,the accuracy of root rot and gray spot images was 83.3%,and the accuracy of other images was above 90%,so after calculating the similarity by the mean hash algorithm,the accuracy of obtaining disease pictures was greatly improved.The data enrichment code was written in Python language,and after several processes such as image rotation,flipping blur,increasing noise,and changing brightness,it reached 17 times amplification,and 2592 soybean disease images were finally obtained.This study improves the accuracy of soybean disease image data collection,and provides a technical reference for the automatic identification and diagnosis of common soybean diseases.