形态学多尺度重建结合凹点匹配分割枸杞图像

针对枸杞分级过程中因图像噪声、光照不均匀和粘连等造成枸杞难以准确分割的问题,提出了一种基于形态学多尺度开闭重建结合凹点匹配的分割方法。首先提取原始图像的红色分量去除枸杞光照阴影噪声,利用形态学多尺度混合开闭重建对红色分量图像进行重建,平滑枸杞内部而保留轮廓边缘信息;然后采用8邻域跟踪算法提取粘连枸杞轮廓边缘;最后运用圆形模板检测粘连枸杞的轮廓凹点,以凹点间最短欧氏距离为匹配条件连接凹点对,并对匹配错误的凹点对进行修正,实现粘连枸杞分割。试验结果表明,该文方法分割准确率较高,而过分割率较低,相比标记控制的分水岭和直接凹点匹配分割等方法,对粘连枸杞分割效果较好,分割准确率可达到96%。该研究可为枸杞分割技术提供理论支撑。

Wolfberry image segmentation based on morphological multi-scale reconstruction and concave points matching

Abstract: The traditional Chinese wolfberry classification usually adopts manual grading in terms of the wolfberry characteristics of size, color, surface defects, and so on. It is a time-consuming and inefficient work. Fortunately, machine vision provides an efficient and fast way to improve the classification efficiency and accuracy. During the process of wolfberry classification by machine vision, the first and important task is to segment wolfberry particles from the image, and then classify them into different grades according to their characteristics. However, the accuracy of wolfberry image segmentation process is often hindered by a number of constraints including noise, inhomogeneous intensity, complex adherent and overlapped particles, which easily cause the decline of segmentation accuracy, and subsequently affect the wolfberry classification effect. For the purpose to improve the accuracy and efficiency of wolfberry image segmentation, a method for efficient segmentation of adherent wolfberries based on morphological multi-scale reconstruction and concave points matching is hereby proposed. Firstly, the red component of the original color image is extracted to partially remove the shadow noise around or inside the wolfberries, and then the red component image is reconstructed by morphological multi-scale mixture opening-closing reconstruction to further smoothen the interior of wolfberries while preserving the contour edge information. Since such reconstruction operation can effectively retain the interesting contour edge of wolfberry particles and eliminate the irregular details, the influence of wolfberries edge contours blur and location offset on the subsequent classification will be greatly reduced. The binary regions of wolfberries are extracted from the reconstructed image by the method of maximum between-cluster variance, and the holes in the interior of wolfberries are filled by morphological filling operator. In the filled binary image, there are 2 kinds of wolfberries. One kind consists of single non-adherent wolfberries particles, and can be extracted by morphological area opening operation without further processing. The other kind mainly contains adherent or overlapped wolfberries particles, and needs to further segment, so 8-neighborhood tracking algorithm is used to extract the edge of single pixel contours of the adherent wolfberries. Taking into account that the shape of wolfberry is ellipsoid, the concave points usually locate in the edges where they are touched or overlapped with each other. Therefore the circular template is used to detect these edge concave points. For the incorrect concave point''s pairs matched by the shortest Euclidean distance as fitting condition, they can be modified according to the unequal pixel distance between the middle point of the connecting line and the boundary point since the length-to-width ratio of the wolfberry is obvious. When all the concave points'' pairs of the adherent wolfberries are confirmed, adherent wolfberries are clearly segmented. The final segmentation results are the combination of single non-adherent and adherent or overlapped wolfberries. The simulation results show that this method can achieve more accurate segmentation results and lower over-segmentation rate compared with the methods of mark-controlled watershed, direct concave points matching, and watershed combined with concave point segmentation, and is especially suitable for the segmentation of adherent wolfberries. The highest accurate segmentation rate is above 96% while over-segmentation rate less than 2%.