随着深度学习技术的迅速发展,卷积神经网络成为研究植物叶部病害识别与病斑分割的主流方法。针对番茄叶部病斑大小不一、形状不规则、病斑分割需要大量像素级标记等问题,文中提出一种多尺度U网络,以同时实现番茄叶部病斑分割与病害识别...随着深度学习技术的迅速发展,卷积神经网络成为研究植物叶部病害识别与病斑分割的主流方法。针对番茄叶部病斑大小不一、形状不规则、病斑分割需要大量像素级标记等问题,文中提出一种多尺度U网络,以同时实现番茄叶部病斑分割与病害识别。在病害特征提取阶段采用多尺度残差模块组合不同尺寸的感受野来提取病害特征,以适应病斑大小和形状的动态变化。引入CB模块(Classifier and Bridge)将病害特征提取阶段与病斑分割阶段连接,对病害特征进行分类,并根据分类结果反向映射出特定类的激活图,此激活图包含特定类别病斑的关键信息。在分割阶段采用上采样与卷积相结合的方法对特定类的激活图进行反卷积,利用跳跃连接方式将反卷积特征与低层特征融合,以补充更多的图像细节信息,获取病斑分割的灰度图。为了使分割的病斑定位更加精确,利用少量像素级标记,对每个像素点采用二分类交叉熵损失函数进行监督训练,同时更好地引导特征提取网络关注病斑部位。利用原始测试集与模拟噪声和光照强度的干扰测试集分别验证模型的病斑分割与病害分类性能。在原始测试样本集上多尺度U网络的平均像素准确率、平均交并比和频权交并比分别达到了86.15%,75.25%和90.27%;在降低30%亮度和添加椒盐噪声的干扰测试集上,模型的识别准确率分别为95.10%和99.20%。实验结果表明,所提方法可以实现番茄叶部病斑分割与识别效果的共同提升。展开更多
As nearly half of the people in the world live on rice, so the rice leaf disease detection is very important for our agricultural sector. Many researchers worked on this problem and they achieved different results acc...As nearly half of the people in the world live on rice, so the rice leaf disease detection is very important for our agricultural sector. Many researchers worked on this problem and they achieved different results according to their applied techniques. In this paper, we applied AlexNet technique to detect the three prevalence rice leaf diseases termed as bacterial blight, brown spot as well as leaf smut and got a remarkable outcome rather than the previous works. AlexNet is a special type of classification technique of deep learning. This paper shows more than 99% accuracy due to adjusting an efficient technique and image augmentation.展开更多
文摘随着深度学习技术的迅速发展,卷积神经网络成为研究植物叶部病害识别与病斑分割的主流方法。针对番茄叶部病斑大小不一、形状不规则、病斑分割需要大量像素级标记等问题,文中提出一种多尺度U网络,以同时实现番茄叶部病斑分割与病害识别。在病害特征提取阶段采用多尺度残差模块组合不同尺寸的感受野来提取病害特征,以适应病斑大小和形状的动态变化。引入CB模块(Classifier and Bridge)将病害特征提取阶段与病斑分割阶段连接,对病害特征进行分类,并根据分类结果反向映射出特定类的激活图,此激活图包含特定类别病斑的关键信息。在分割阶段采用上采样与卷积相结合的方法对特定类的激活图进行反卷积,利用跳跃连接方式将反卷积特征与低层特征融合,以补充更多的图像细节信息,获取病斑分割的灰度图。为了使分割的病斑定位更加精确,利用少量像素级标记,对每个像素点采用二分类交叉熵损失函数进行监督训练,同时更好地引导特征提取网络关注病斑部位。利用原始测试集与模拟噪声和光照强度的干扰测试集分别验证模型的病斑分割与病害分类性能。在原始测试样本集上多尺度U网络的平均像素准确率、平均交并比和频权交并比分别达到了86.15%,75.25%和90.27%;在降低30%亮度和添加椒盐噪声的干扰测试集上,模型的识别准确率分别为95.10%和99.20%。实验结果表明,所提方法可以实现番茄叶部病斑分割与识别效果的共同提升。
文摘As nearly half of the people in the world live on rice, so the rice leaf disease detection is very important for our agricultural sector. Many researchers worked on this problem and they achieved different results according to their applied techniques. In this paper, we applied AlexNet technique to detect the three prevalence rice leaf diseases termed as bacterial blight, brown spot as well as leaf smut and got a remarkable outcome rather than the previous works. AlexNet is a special type of classification technique of deep learning. This paper shows more than 99% accuracy due to adjusting an efficient technique and image augmentation.
