融合K-means与Ncut算法的无遮挡双重叠苹果目标分割与重建

重叠苹果目标的准确分割是采摘机器人必须解决的关键问题之一。针对现有重叠苹果目标分割方法不能保留重叠部分轮廓信息的问题,提出了一种无枝叶遮挡的双果重叠苹果目标分割方法。该方法首先利用K-means聚类算法进行图像分割以提取苹果目标区域,然后利用Normalized Cut(Ncut)算法提取苹果目标轮廓,以实现未被遮挡苹果目标完整轮廓的准确提取,最后利用Spline插值算法对遮挡的苹果目标进行轮廓重建。为了验证算法的有效性,对20幅无枝叶遮挡双果重叠的苹果图像进行试验,并将该算法与寻找2个有效凹点用其连线分割重叠苹果目标,把分离的2个轮廓分别用Hough变换重建轮廓的方法进行对比。试验结果表明,对于图像中未被遮挡的苹果目标,利用该研究算法的平均分割误差为3.15%,提取的苹果目标与原始图像中苹果目标的平均重合度为96.08%,平均误差比Hough变换重建算法低7.73%,平均重合度高9.71%,并且该研究算法能够很好地保留未被遮挡苹果目标的完整轮廓信息,提高了分割精度。对于重叠被遮挡的苹果目标,平均分割误差和平均重合度分别为5.24%和93.81%,比Hough变换重建算法的平均分割误差低11.35%,平均重合度高12.74%,表明该算法可以较好地实现重叠被遮挡苹果目标的轮廓重建,研究结果可为实现枝叶遮挡影响下的多果重叠目标分割与重建提供参考。

Fusion of K-means and Ncut algorithm to realize segmentation and reconstruction of two overlapped apples without blocking by branches and leaves

Overlapped apples are common in natural scenes, and they can seriously affect the execution of picking task. Accurate segmentation of overlapped apples is one of the key problems that picking robot must solve. As the existing overlapped apples segmentation methods could not retain the contour of overlapped parts, a new method of segmenting overlapped apples was proposed in this paper. The steps of the algorithm were as follows: Firstly, the image was processed with morphological opening operation using disk-shaped structural element with the radius of 5 pixels so as to ensure the data information consistent in small area and to make the image more smooth. In order to distinguish the target and the background automatically, the image was then transformed from RGB color space into L*a*b color space. Then K-means clustering method was used to segment images to extract apple region. Secondly, in order to extract accurate and entire contour of apple region and keep the contour of overlapped part simultaneously, Normalized Cut (Ncut) method was adopted. To ensure the accuracy of Ncut algorithm, the cluster number was chosen as 5. Thirdly, contour segmentation method was used to extract contour of each apple. The procedures of contour segmentation method were as follows: 1) The extracted contours were connected by simply using eroding operation in order to avoid disconnection at intersection of contours of 2 adjacent apples; 2) Refine the connected contour and search three cross point, then break off the contour at three cross point by assigning zero at three cross point, and 3 contours could be got; 3) Add the 2 contours whose bending directions were opposite, and this was the contour of unblocked apple and the other contour was the contour of blocked apple. Lastly, Spline interpolation method was used to reconstruct the contour of blocked apples. In order to verify the validity of this algorithm, 20 images of adjacent apples were used to conduct the experiment, and the result was compared with the comparison method. The comparison method is a method that uses the connection line of 2 concave points to segment overlapped apples, and then utilizes the Hough transform method to reconstruct the contour of apples. The experimental results showed that for unblocked apples in apple images, average segmentation error of the presented method was 3.15%, 7.73% less than that of comparison method (10.88%). Average overlap ratio of the presented method was 96.08% and was increased by 9.71% compared to comparison method (90.85%). In addition, this method could keep the complete contour information of unblocked apples and thus improved segmentation accuracy effectively. For blocked apples, average segmentation error and average overlap ratio were 5.24% and 93.81%, respectively. The segmentation error was decreased by 11.35% and average overlap ratio was increased by 12.74% compared to comparison method, which indicated that the method could reconstruct contour of blocked apples well. In conclusion, the presented algorithm is feasible to segment and reconstruct 2 overlapped apples without blocking by branches and leaves. However, for the images of overlapped apple blocked by branches and leaves, the images of more than 2 overlapped apples, and the images of 2 overlapped apples unblocked by branches and leaves with unclear contour at overlapped part, this method cannot complete the segmentation, and thus further research would be needed.