基于全卷积网络的葡萄病害叶片分割

[目的]本文旨在解决不同光照和复杂背景下葡萄病害叶片图像的自动分割。[方法]使用了一种全卷积网络（FCN）的葡萄病害叶片图像的自动分割算法。该算法在结构上将传统的卷积神经网络（CNN）后3个全连接层换成3个卷积层。通过多层的卷积，对输入葡萄叶片图像的特征进行提取；通过池化层，对特征信息进行筛选，缩减特征尺寸，以达到减少网络参数的目的。再通过反卷积对特征上采样，从高维、小尺寸特征恢复到图像原始尺寸，对具有原始尺寸的特征进行逐像素分类，确定原图像中每个像素位置的标签是背景还是前景。因只经过上采样处理后的分割图像会较粗糙，故通过跳跃结构将较为粗糙的原图进行局部信息与整体信息的整合，达到对分割结果进行精细化处理的目的。[结果]本算法对葡萄病害叶片有较好的分割效果，单叶片和复杂多叶片图像的马修斯相互系数（MCC）分别为0.821和0.747，MCC平均值较对比算法提高了6.5%。[结论]本算法能够较精确地分割自然条件下成像的葡萄病害叶片图像，为后续在叶片精准分割病害区域和提取病害特征创造了良好的条件。
[Objectives]The research aimed to solve the automatic segmentation of diseased grape leaf images under different light and cluttered background.[Methods]We used a full convolution network(FCN)to automatically segment the grape leaf images. The method replaced the last three full connection layers in a tradition convolution neural network(CNN)with three convolution layers. Through multiple convolution layers,the features of input of images were extracted. And by the pooling layers,feature sizes were reduced so the network parameters decreased. When the features were gotten,they were up-sampled with the de-convolution layers,restoring the original image size. The output was labeled background and foreground with pixel wise classifiers. However,the segmentation was kind of rough. So,two skipping structures were employed to get finer results by integrating the local information and whole information.[Results]The experimental results showed that the algorithm worked well in segmenting the grape leaves of diseases. The mathews mutual coefficient(MCC)achieved 0.821 and 0.747 for single leaf and multiple leaves,respectively. The average MCC improved 6.5% than the contrasted algorithm.[Conclusions]Due to the good performance of the proposed algorithm,it can create good conditions for the subsequent segmentation of the disease area in leaves and the extraction of disease features