1. 山东师范大学信息科学与工程学院, 山东 济南 250358
2. 山东师范大学分布式计算机软件新技术山东省重点实验室, 山东 济南 250358
3. 山东理工大学农业工程与食品科学学院, 山东 淄博 255000
4. School of Engineering, Cardiff University, Cardiff, CF24 3AA, United Kingdom
KDC: A Green Apple Segmentation Method
WANG Zhi-fen1, JIA Wei-kuan1,2*, MOU Shan-hao1, HOU Su-juan1,2, YIN Xiang3, ZE Ji4
1. School of Information Science and Engineering, Shandong Normal University, Ji’nan 250358, China
2. Shandong Provincial Key Laboratory for Novel Distributed Computer Software Technology, Shandong Normal University, Ji’nan 250358, China
3. School of Agricultural Engineering and Food Science, Shandong University of Technology, Zibo 255000, China
4. School of Engineering, Cardiff University, Cardiff, CF24 3AA, United Kingdom
摘要： 苹果的可见光谱目标的高效、精准识别是实现果园测产或机器自动采摘作业的关键，由于绿色目标果实与枝叶背景颜色较为相近，因此绿色苹果的识别成为新的挑战。再由于果园实际复杂环境因素影响，如光照、阴雨、枝叶遮挡、目标重叠等情况，现有的目标果实识别方案难以满足测产或自动采摘的实时、精准作业需求。为更好地实现果园自然环境中绿色目标果实识别问题，提出一种新的核密度估计优化的聚类分割算法（kernel density clustering, KDC）。新算法首先利用简单的迭代聚类（simple linear iterative cluster, SLIC）算法将目标图像分割成不规则块，集结小区域内近似像素点组成超像素区域，计算单元由像素点转变为超像素区域，有效降低数据复杂度，且SLIC算法简化图像数据时可有效避免目标果实轮廓模糊；基于超像素构造R-B区域均值和G-B区域均值的二维特征分量，建立针对聚类分析的青苹果颜色特征空间。然后借助密度峰值聚类中心计算绿色苹果图像每个数据点的局部密度和局部差异度，为解决分割边界模糊问题，在计算过程中利用核密度估计计算局部密度，确保局部密度在不同复杂场景中的清晰准确表达，以更精准找出被低密度区域分割的高密度区域，实现任意形状的聚类。最后以局部密度和距离构造寻找聚类中心的决策图，该研究采用双排序算法实现聚类中心的自动选择，完成目标果实的高效分割。新算法通过SLIC算法获得图像的超像素区域表示，数据点的局部密度通过核密度估计得到，大幅降低算法的计算量，实现目标图像的高效、精准分割。为更好地验证新算法性能，实验采集多光照、阴雨等环境下的遮挡、重叠等复杂目标图像，以分割效率、分割有效性、假阳性、假阴性等指标进行评价，通过对比k-means聚类算法、meanshift聚类算法、FCM算法和DPCA算法，该研究提出的新算法分割性能均最优。
关键词：绿色果实；图像分割；密度峰值聚类；核密度估计
Abstract：Efficient and accurate recognition of apples in the visible spectrum is the key part of realizing orchard production prediction and automatic harvesting. The color of the green target is similar to the background of the branches and leaves.Thus the recognition of green apples becomes a new challenge. In the natural orchard, many complicated environmental factors may affect the accurate recognition of target fruit, such as light, rain, occlusion of branches and leaves, and overlap of targets. The existing recognition methods of target fruit are difficult to realize the real-time and accurate production prediction and automatic harvesting. In order to better realize the recognition of the green target in the natural orchard, a new kernel density clustering (KDC) is proposed. Firstly, the target image is segmented into irregular blocks by the simple linear iterative cluster (SLIC) algorithm, and approximate pixels in a small area are merged into superpixel area. The basic calculation unit is converted from a pixel point to a superpixel area to reduce the complexity. The SLIC algorithm is used to effectively avoid the target fruit contour’s blur when simplifying the image data. Based on superpixels, the two-dimensional feature components of the R-B region mean and the G-B region means are constructed, and the color feature space of the green apples for cluster analysis is established. Then, the density peak clustering algorithm calculates the local density and local difference degree of each data point of the green apple image. The preciseand accurate local density in different complicated scenes is calculated using kernel density estimation. The high-density area segmented by the low-density area is found more accurately, and the clustering of arbitrary shapes is realized. Finally, a decision graph used to find a cluster center is constructed according to local density and distance. A double sorting algorithm is employed to automatically select cluster centers and efficient segmentation of target fruit. The superpixel area representation of the image is obtained using the SLIC algorithm, and the local density of the data points is obtained using kernel density estimation. The new algorithm greatly reduces the computational complexity and achieves efficient and accurate segmentation of the target images. In order to better verify the performance of the new algorithm, the experiment collects occluded and overlapping target images under multiple lighting, rainy and other complex environments and evaluates them based on segmentation efficiency, segmentation effectiveness, false positives, and false negatives. By comparing the k-means clustering algorithm, mean shift clustering algorithm, fuzzy C-means (FCM) algorithm and density peak clustering algorithm (DPCA), it is finally concluded that the segmentation performance of the new algorithm is the best.
Key words：Green fruit; Image segmentation; Kernel density estimation; Density peak clustering
基金资助: Focus on Research and Development Plan in Shandong Province (2019GNC106115)，China Postdoctoral Science Foundation (2018M630797), Shandong Province Higher Educational Science and Technology Program (J18KA308), National Nature Science Foundation of China (62072289)
通讯作者:
贾伟宽
E-mail: wkjia@sdnu.edu.cn
作者简介: WANG Zhi-fen，(1998—)，Master student, School of Information Science and Engineering, Shandong Normal University
e-mail:
xingchen0612@gmail.com
引用本文:
王志芬, 贾伟宽, 牟善昊, 侯素娟, 印 祥, ZE Ji. 基于核优化密度聚类的绿色苹果分割算法[J]. 光谱学与光谱分析, 2021, 41(09): 2980-2988.
WANG Zhi-fen, JIA Wei-kuan, MOU Shan-hao, HOU Su-juan, YIN Xiang, ZE Ji. KDC: A Green Apple Segmentation Method. SPECTROSCOPY AND SPECTRAL ANALYSIS, 2021, 41(09): 2980-2988.
[1] Jia W K, Zhang Y, Lian J, et al. International Journal of Advanced Robotic Systems, 2020, 17(5-6): 25310.
[2] Gongal A, Karkee M, Amatya S. Information Processing in Agriculture, 2018, 5(4): 498.
[3] Lv J, Wang Y, Xu L, et al. Scientia Horticulturae, 2019, 257: 108758.
[4] LI Da-hua, ZHAO Hui, YU Xiao. Spectroscopy and Spectral Analysis, 2019, 39(9): 2974.
[5] Jia W, Tian Y, Luo R, et al. Computers and Electronics in Agriculture, 2020, 172: 105380.
[6] Bulanon D M, Burks T F, Alchanatis V. Biosystems Engineering, 2009, 103(1): 12.
[7] Li B, Long Y, Song H. International Journal of Agricultural and Biological Engineering, 2018, 11(1): 192.
[8] Maldonado Jr W, Barbosa J C. Computers and Electronics in Agriculture, 2016, 127: 572.
[9] Huang X Y, Li GL, Ma C, et al. Transactions of the Chinese Society of Agricultural Engineering, 2018, 34(23): 142-148. (in Chinese)
[10] Wang D D, He D J, Song H B, et al. Multimedia Tools and Applications, 2019, 78(13): 17391.
[11] Achanta R, Shaji A, Smith K, et al. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2012, 34(11): 2274.
[12] Liu R, Wang H, Yu XM. Information Sciences, 2018, 450: 200.
[13] Tian K, Li J, Zeng J, et al. Computers and Electronics in Agriculture, 2019, 165: 104962.
[14] Guo Y, Şengür A, Akbulut Y, et al. Measurement, 2018, 119: 28.
[15] Lei T, Jia X, Zhang Y, et al. IEEE Transactions on Fuzzy Systems, 2018, 27(9): 1753.
