| 利用特征分割和病斑增强的杨树叶部病害识别 |
| |
| 引用本文： | 明浩,苏喜友. 利用特征分割和病斑增强的杨树叶部病害识别[J]. 浙江农林大学学报, 2020, 37(6): 1159-1166. DOI: 10.11833/j.issn.2095-0756.20190752 |
| |
| 作者姓名： | 明浩 苏喜友 |
| |
| 作者单位： | 北京林业大学 信息学院，北京 100083 |
| |
| 基金项目： | “十三五”国家重点研发计划项目(2017YFD0600906) |
| |
| 摘 要： | 目的 针对杨树Populus黑星病早期特征和杨树花叶病病斑不明显的特点，提出通过对图像集进行预处理的方法以提高识别精度的方案。 方法 为去除图像背景的影响，采用基于改进的Canny算子边缘检测法并结合霍斯变换提取叶片轮廓；借助限制对比度自适应直方图均衡化算法降低局部光照不均带来的影响并增强病斑的特征；使用自适应阈值的OTSU分割算法提取病斑图像。最后将预处理得到的病斑特征二值化图像和病斑图像，分别输入由5个卷积层、3个全连接层、650 000个神经元及超过6 000万个学习参数的Alexnet神经网络进行训练并验证准确率。 结果 研究最终分别获得93.56%和98.07%的验证集识别精度，较原图像实验组88.77%的识别精度有显著提升。提出的提取叶片轮廓的结合方法能够完整提取不同背景下的叶片主体图像，有效避免目标叶片的背景干扰；限制对比度自适应直方图均衡化算法对自然环境下拍摄产生的不均匀光照有较好的处理效果，有效降低反光等因素的干扰。 结论 几种病害图像预处理对提高识别精度效果明显，识别能力远超过未经处理的原始病害图像识别，有助于提高杨树叶部病害的智能识别能力。图8表1参22
|
| 关 键 词： | 森林保护学 杨树病害 卷积神经网络 图像分割 病斑识别 |
| 收稿时间： | 2019-12-24 |
Image recognition of poplar leaf diseases with feature segmentation and lesion enhancement |
| |
| MING Hao,SU Xiyou. Image recognition of poplar leaf diseases with feature segmentation and lesion enhancement[J]. Journal of Zhejiang A&F University, 2020, 37(6): 1159-1166. DOI: 10.11833/j.issn.2095-0756.20190752 |
| |
| Authors: | MING Hao SU Xiyou |
| |
| Affiliation: | School of Information Science & Technology, Beijing Forestry University, Beijing 100083, China |
| |
| Abstract: | Objective In view of the inconspicuous characteristics of poplar(Populus) scab and mosaic disease, this paper is aimed to propose a method to improve the disease recognition accuracy by means of the pre-treatment of the original image set. Method Firstly, the contour of the blade was extracted employing the improved Canny operator edge detection method combined with Hoss Transformation so as to remove the disturbance of the image background. After that, the contrast limited adaptive histogram equalization was adopted to reduce the impact of local illumination unevenness. Thirdly, the OTSU segmentation algorithm with adaptive threshold was used to extract leaf lesion images. At last, the binarized images with lesion and the leaf lesion images were fed into the Alexnet which consists of 5 convolutional layers, 3 full-connection layers, 650 000 neurons and over 60 million learning parameters. Result Both groups came back with a significantly higher recognition accuracy rate than that of the the original image experiment group (93.56% and 98.05% VS 88.77%). The hybrid method proposed in this paper could help completely extract the images of the main body of the blade with different backgrounds and effectively avoid the background interference of the target blade. And the adaptive histogram equalization algorithm with limited contrast helped in dealing with the uneven light produced by natural environment and reducing the interference of reflective factors like reflected light. Conclusion The pre-treatment of the images of the above-mentioned diseases has significantly improved the recognition accuracy, and is highly recommended in future tasks. [Ch, 8 fig. 1 tab. 22 ref.] |
| |
| Keywords: | |
|
| 点击此处可从《浙江农林大学学报》浏览原始摘要信息 |
|
点击此处可从《浙江农林大学学报》下载免费的PDF全文 |
