基于融合多特征图切割的作物病害图像自动分割

为提高黄瓜叶部病害图像的分割性能,该文提出一种基于融合多特征图切割的病害图像自动分割方法。首先采用一种新的阈值化方法对原始病害图像的红色分量进行二值化处理;然后融合纹理、灰度、距离3个特征构建能量函数的边界项,描述像素间的相似性;再利用分割区域像素与区域边界像素的红色分量差值自动建立能量函数的区域项,反映像素归属于背景和目标的程度;最后运用最大流算法求解能量函数得到分割结果。将该方法应用于黄瓜3种病害(靶斑病、霜霉病和白粉病)叶部图像分割中,并与OTSU算法及半自动图切割算法的分割结果进行比较。试验结果表明,该方法的平均错分率为1.81%,低于其他2种算法,平均分割速度约为2.34 s并无大幅增加。该研究可为黄瓜病害的自动识别和诊断提供技术参考。

Automatic segmentation of plant disease images based on graph cuts fusing multiple features

Abstract: Diseases in crops can lead to declines of production and quality, which cause economic losses in agricultural industry worldwide. Therefore, detection of the diseases in plants is extremely critical for sustainable agriculture. Many crop diseases perform on the leaves, and often present in the form of spots, so processing the leaf images is a feasible way for identifying and diagnosing diseases. Spots separation from the leaf is a very important step in the process of disease recognition and diagnosis. And the segmentation accuracy affects the reliability of the subsequent feature extraction and the accurateness of following classification directly. To improve the segmentation performance, an automatic segmentation algorithm based on graph cut which fused multiple features was put forward in this paper. Firstly the background was excluded by threshold method so as to speed up the image segmentation. The experimental comparison results showed that the segmentation effect for color images processed by OTSU algorithm, a simple adaptive threshold image segmentation method, was not very satisfactory. Therefore, a new method of threshold processing was studied here, which could remove most of the background and did not lose the disease spots. In addition, the red component of the original color images was used as the research objects of the new threshold method, since it had the strongest contrast compared with the green component and the blue component. Then three features, texture, gray level and distance were fused to build the boundary term of the energy function, which described the similarity between the pixels. Among them, for the sake of reducing the computational complexity and calculation time, the texture feature was simply defined by the one-dimensional entropy of images, which was the amount of information included by the gathered characteristics of gray level distribution. Moreover, in order to reflect the extent of the pixel belongs to the background or target, the red component difference between pixels in the image region and the region boundary was used to set up the area term of the energy function automatically. Finally the maximum flow algorithm was utilized to solve the established energy function, and the segmentation results were obtained. With the purpose of verifying the validity of the proposed algorithm, the method was applied to divide three kinds of cucumber disease (target spot, downy mildew and powdery mildew) leaf images. Each disease of 50 pictures, a total of 150 pictures were selected randomly as the experimental samples. And the OTSU algorithm and semi-automatic graph cuts algorithm were chosen as the contrast means. The experimental results demonstrate that the disease spots of the leaf images can be separated effectively when using the method proposed in this paper. The average error rate is 1.81%, which is lower than the other two algorithms, and its average segmentation time do not significantly increase. This study can provide a technical reference for the automatic identification and diagnosis of cucumber diseases in the future.