吸引权重加权FCM的土壤图像阴影检测  

Shadow Detection of Soil Image Based on Attraction-Weighted FCM

作　　者：佘春燕 曾绍华[1,2] 王帅 徐毅丹 SHE Chunyan;ZENG Shaohua;WANG Shuai;XU Yidan(College of Computer and Information Science,Chongqing Normal University;Chongqing Center of Engineering Technology Researchon Digital Agricultural Service,Chongqing 401331;Chongqing Master Station of Agricultural Technology Promotion,Chongqing 400121;Chongqing Beibei District Station of Plant Protection Plant Quarantine Station,Chongqing 400715,China)

机构地区：[1]重庆师范大学计算机与信息科学学院 [2]重庆师范大学重庆市数字农业服务工程技术研究中心,重庆401331 [3]重庆市农业技术推广总站,重庆400121 [4]重庆市北碚区植物保护植物检疫站,重庆400715

出　　处：《重庆师范大学学报（自然科学版）》2020年第6期114-125,F0002,共13页Journal of Chongqing Normal University:Natural Science

基　　金：重庆市教育委员会科学技术研究重点项目(No.KJZD-K201900505);重庆市高校创新研究群体(No.CXQT20015);重庆师范大学科研项目(No.YKC20032)。

摘　　要：【目的】自然环境下机器视觉采集的土壤图像存在阴影,而阴影检测是后续亮度归一化和土种识别的重要预处理工作,基于此提出算法对机器采集的土壤图像进行阴影检测。【方法】首先,对图像的L分量和I分量的密度峰值集进行优化截断对齐,通过高斯平滑获取主峰值点,得到基于L分量和I分量的2因素密度峰值,作为改进模糊C均值(FCM)的自适应初始聚类中心;然后,引入拉伸因子对数据进行拉伸,提升数据差异;最后,定义基于全局密度和类面积的吸引权重,重构FCM优化模型,实现基于FCM的土壤图像自适应阴影检测。【结果】提出的算法检测的阴影区域和非阴影区域的平均亮度标准差分别为25.988 0,27.981 4,比对比算法降低了1.04%~32.23%;提出的算法平均迭代次数和平均运行时间分别为3次和1.515 8s。【结论】提出的算法可实现自然环境下土壤图像阴影检测,具有检测精度高、时间花销小的特点。研究结果为土壤图像阴影检测提供了参考。[Purposes]Due to the unevenness of the soil surface or the occlusion of the shooting environment on the spot,the soil image obtained by machine vision in natural environment has discrete and large area shadow.In order to avoid the interference of shadow on further recognition of soil with machine vision,it is a significant pre-processing task to detect accurately shadows for subsequent illumination normalization and soil recognition.[Methods]Aiming at the problem of FCM falling into local optimum due to the selection of initial parameters,the algorithm of initializing adaptively the clustering centers was proposed here.The specific method was as follows:all the peaks of the histogram which the Land I components of the image were obtained,then the truncation alignment model of the peak point was established with the peaks.This optimization model was used to select two sub-peaks sets,which were derived from the Land I component histograms,respectively.Extracting the frequency corresponding to the sub-peaks in the original histogram to form a sub-histogram,and then it was repeatedly smoothed by Gaussian until remaining two peaks.The remaining two peaks were the two-factor density peaks,which were used as adaptive initial clustering center of the improved FCM.As we all know,FCM is a clustering algorithm to minimize the loss function by iteratively optimizing the clustering center and membership matrix.To improve the samples difference,a stretching factor was introduced to stretch the samples.The degree of stretching depends on the spatial position relationship between the sample and the clustering center.Aiming at the shortcoming that the FCM didn’t consider the distribution of data,it introduced a new weighted FCM algorithm based on global density and the area of the cluster.It reflected the distribution of data through the density and area in the two-dimensional plane,which can be integrated into the FCM model by weighting the distance between the data point and the cluster center.Then,our algorithm was applied t

关 键 词：阴影检测 FCM 土壤 机器视觉 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]