基于显微图像处理的稻瘟病菌孢子自动检测与计数方法

稻瘟病菌孢子的检测通常在显微镜下由人工目测完成,该方法费时、费力、自动化程度低。因此,该研究提出了一种基于显微图像处理技术的稻瘟病菌孢子自动检测和计数方法。首先,采用显微图像系统获取稻瘟病菌孢子图像;然后提出一种分块背景提取法对其进行光照校正;根据显微图像中孢子的边缘特征,利用Canny算子进行边缘检测,其中Canny边缘检测过程中的阈值应用模糊C均值算法在梯度图上自动确定;接着对边缘检测后的二值图像进行数学形态学闭开运算处理。根据孢子和主要杂质的形态特征,利用椭圆度、复杂度和最小外接矩形宽度等形态特征参数对目标物进行分类,提取只含孢子的二值图像。最后,提出了基于距离变换和高斯滤波的改进分水岭算法对粘连孢子进行分离。测试结果表明:在100幅测试的显微图像样本中,孢子检测的平均准确率为98.5%,满足稻瘟病菌孢子自动检测和计数要求。

Automatic detection and counting method for spores of rice blast based on micro image processing

Abstract: The detection and counting for spores of the rice blast usually relies on the eye observation under a microscope, which is time consuming, labor intensive and inefficient, so an alternative method is required. This paper discussed an innovative method using image processing techniques to detect and count the spores in micro images. Firstly, the micro images of spores were captured with image detection system consisting of a microscope, a video camera, a capturing software and a computer. And then, a correction method was presented to reduce the non-uniform illumination by subtracting the gray value of the background image form the original image. The original image was divided to 4×4 blocks and the gray value of the background was determined by the illumination correction for each dividing part. The spores had strong edge information in the micro images, so the canny operation was applied to do the edge detection. In this process, fuzzy c-means algorithm (FCM) was used to obtain the high threshold of the canny operation automatically in the gradient images. The noises especially for mycelium could be filtered better using FCM-Canny than Ostu-Canny method. Morphological image processing including close and open operations was implemented to fill the spores and filter the noises. According to the differences of the shape characteristics between the spores and the other objects, the features' combination composed of ellipticity, complexity and width of minimum bounding rectangle was selected after sampling statistics to recognize the spores. When 0.85< ellipticity<1.33, complexity <2.1 and width of minimum bounding rectangle >20, the objects were recognized as the spores, otherwise deleted as noises. The binary images including only spores were gained by a series of image processing, but there were still some adjacent spores in the images. In order to count the spores precisely, these adjacent ones must be separated. This paper presented an improved watershed algorithm (WA) to break the adjacent parts for getting the right number. The binary images of the spores were transformed to the gray images by distance transform (DT), then Gaussian filtering (GF) was applied to unite the redundant local minimum for preventing the over segmentation, and the WA was conducted to separate the adjacent spores at last. To verify the the proposed method, a total of 100 images were collected for the performance evaluation. Experimental results showed that the numbers of the image samples were 79 with detection accuracy of 100%, 16 with detection accuracy from 90% to 100% and 5 with detection accuracy from 80% to 90%. The proposed method achieved high-accuracy detection and counting with average accuracy of 98.5%, which met the requirements of the automatic detection and counting for spores of the rice blast.