基于离散余弦变换和区域生长的白粉虱图像分割算法

图像分割是病虫害自动识别的难点之一,目前大多基于颜色、纹理等信息采用阈值法或聚类法进行分割,简单,易实现,但分割精度较低。该文针对田间开放环境中,不能用颜色、纹理特征有效分割病虫害图像的问题,引入离散余弦变换(discrete cosine transform,DCT),提出用清晰度对病虫害图像进行分割,以提高分割精度。DCT的低频信号表示图像轮廓,高频信号表示图像细节,对于病虫害图像,焦点通常聚集在目标区域,该文提出截断DCT高频信号,再与原图做差的方法以区分清晰部分和模糊部分,然后结合病虫图像局部聚合度较高的特性,利用区域生长方法提取完整目标。采用该算法对白粉虱图像进行分割测试,并与阈值法和GMM方法比较:分割结果中,目标的一致性和边缘的清晰度明显好于阈值法和GMM方法,平均正确分类率为98.49%,分别较R,B,Y空间中阈值法和Y空间中GMM方法分类正确率高2.96%、3.28%、3.24%和9.65%,差异达到显著水平。基于DCT和区域生长的分割算法鲁棒性高,能够有效地将病虫害区域从自然环境中采集的叶片中分离,可用于分割白粉虱图像。

Algorithm for segmentation of whitefly images based on DCT and region growing

Image segmentation is one of the fundamental problems in an automatic pest identification system. In the current research, algorithms based on thresholding or clustering are widely used. Despite the simplicity and efficiency of the traditional methods, their performances are not satisfactory because the gray intensity is overlapped among the background of plant leaves and pests in the field environment. In this paper, we propose a novel method to segment the whitefly in the field environment by the Discrete Cosine Transformation (DCT) and region growing methods. The images are assumed to be rightly taken and focused on the target objects. The low frequency of DCT represents the image contour, and the high frequency of DCT represents the image details. The high frequency of DCT is used to distinguish the blurred image from the clear image globally. On the other hand, the local intensity of the pests is changed gradually and the intensity between pests and the closed background or plant leaves is changed greatly, so region growing is adopted to take advantage of the local intensity of the objects and to extract complete targets locally. To be specific, first, the gray image is transformed by discrete cosine transformation, and the high frequency part is truncated. Then it is re-converted to a gray image by inverse discrete cosine transformation. Second, the transformed image and original image are differentiated. Through an adaptive thresholding and open-close operation, we obtained the clear foreground regions. Third, we marked each clear region and established the gray model. Finally, as the pests have good local polymerization degree, the region growing method was adopted to extract the complete target object. Pixels in the clear regions and conforming the region gray model are involved in the growing process with an 8-direction searching scale. As a result, each single connected component was taken as a target pest. The algorithm was implemented on a Visual Studio 2005 platform. The experiments were conducted on whitefly images by comparison with the methods based on thresholding and Gaussian Mixture Model (GMM). The average classification accuracy was 98.49%, which was higher than thresholding-based methods in space R, B, Y and GMM in space Y, respectively, by 2.96%, 3.28%, 3.24% and 9.65%. Experimental results show that our proposed method can effectively separate pests apart from normal part of leaves and background. Our method provides higher precision as well as the accurate and closed boundaries, which is beneficial in the processing of whitefly images.