采用全景技术的机器视觉测量玉米果穗考种指标

Investigate maize ear traits using machine vision with panoramic photography

为了在利用图像技术无损考察玉米果穗形态指标时,能够利用一幅图像显示整个玉米果穗的外形,从而减少多幅图像拼接产生的重叠和处理不便,该文提出一种新的基于机器视觉的玉米果穗考种方法与配套装置,首先拍摄旋转玉米果穗图像序列,应用SIFT(scale invariant feature transform)算法获取图像特征点,对特征点随机采样计算单应矩阵并进行一致性检测排除外点,将前后2帧图像注册到同一坐标系。然后采用动态规划法寻找前后2帧拼接图像的缝合线,按缝合线切割图像,以图像模板高斯滤波权值融合缝合线两侧图像消除曝光差异。依次拼接、融合图像序列生成果穗全景图。对果穗全景图进行考种指标检测,试验结果表明:基于机器视觉的测量值与人工测量方式不存在显著性差异(显著水平α=0.05),该文所述方法可满足自动化考种的需求。

Maize ear morphological characteristics have important applications in breeding, germplasm, and cultivation areas, subject to the extent of technology development in relevant areas, but the approach of surveying morphological characteristics is not highly automatic. In this paper, we present a new machine vision based method and a supporting device for maize ear morphological characteristic surveying. First, the maize ear was placed on a rotating component, which rotates the maize ear in a fixed angle interval in order to capture 16 images more or less. A preprocess was carried out of maize ear image sequences to remove the image background, and the remaining part of the maize ear image was passed to the next process. The SIFT （Scale Invariant Feature Transform） was used to extract image feature points, and the feature points in the neighboring images could be matched up according to SIFT feature points. The relative motion between the two images could be described by a homography, and an overdetermined equations composed of matching points and homography make specific values of homography available. Mismatched feature points will reduce the accuracy of the homography equation solution dramatically. We adopted a RANSAC （random sample consensus） method to remove the outlier of the matching points during the homography solving process. Secondly, according to the motion described by homography, the first image and the next image are registered to the same coordinate system, using the dynamic programming method to find the seam-line in the two images, cutting the redundancy region in the two images along the seam-line. Since the exposure of the two images had certain differences which led to image brightness near seam-line being slightly different, a weighted Gaussian filter was imposed on both sides of the stitching image to eliminate exposure difference. Finally, the fusion image according to the order in sequence generated the ear panorama, row number, number in a row, kernel number, and other parameters were extracted by processing the maize ear panorama. The experimental results showed that： there is no significant difference between the method proposed by this paper and manual measurement, and the method proposed can greatly strengthen the automation of the maize ear traits investigation.