多尺度U网络实现番茄叶部病斑分割与识别

Multi-scale U Network Realizes Segmentation and Recognition of Tomato Leaf Disease

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。

With the development of deep learning technology,convolutional neural network has been the mainstream method for plant leaf disease recognition and disease spot segmentation.Aiming at the problems of different sizes and irregular shapes of tomato leaf lesions,need for a large number of pixel-level labels,a novel multi-scale U network is proposed,which realizes tomato leaf lesion segmentation and disease recognition simultaneously.For disease feature extraction,a multi-scale residual module including different sizes of receptive fields is used to extract disease features according to the different disease spot size and shape.The CB module(Classifier and Bridge)is introduced to connect the disease feature extraction stage with the lesion segmentation stage,which classifies the disease and also reversely generates an activation map of specific class according to the classification result.This activation map contains the specific type of lesions label information.In the segmentation stage,upsampling and convolution are used to deconvolve the activation map.The deconvolution feature and the low-level feature are merged by the jump connection method.In order to make lesion location segmentation more accurate,a few of pixel-level labels are used for training to minimize two-class cross-entropy loss.In the experiment,the original samples and samples with simulated noise and light intensity are used to verify the performance of disease spot segmentation and disease recognition of our method.On the original sample set,the average pixel accuracy,average intersection ratio,and frequency weight intersection ratio of our method reaches 86.15%,75.25%,and 90.27%,respectively.In the interference sample with 30%brightness reduction,salt and pepper noise,the recognition accuracy of our method obtains 95.10%and 99.20%respectively.Experimental results show that the proposed method can achieve improvement in segmentation and recognition of tomato leaf lesions simultaneously.