基于半监督离散度的土壤彩色图像阴影检测

Shadow Detection of Soil Color Image Based on Semi-Supervised Dispersion

【目的】由于光照及土壤心土自然断口凹凸的影响,机器视觉采集终端获取的土壤图像中存在阴影,为避免对后续土种识别造成干扰,研究对土壤图像进行阴影检测的方法。【方法】通过对土壤彩色图像HSI颜色空间阴影与非阴影分析,发现阴影与非阴影分别在色调(H)和亮度(I)分量具有一定的分离特性;首先,为了增大分离特性用于土壤图像阴影检测,将H转换为H″,并利用H″和I引入拉伸因子构建具有显著双峰和分离特性的m测度,用于阴影检测;然后,利用m测度直方图双峰特性粗略估计阴影检测阈值,并根据粗略估计的阴影检测阈值及2个主峰点,对部分数据做阴影与非阴影标定,分析获取阴影与非阴影区域监督信息;最后,构建待检测数据子集和定义它与阴影和非阴影监督信息的离散度,逐步对未标定数据进行半监督聚类,完成土壤彩色图像阴影检测。【结果】本文算法分割土壤图像非阴影和阴影标准差分别为0.063,0.058,检测的土壤图像非阴影和阴影标准差非常接近且数值较小,说明算法是有效的,和已有文献中的算法相比本文检测的土壤图像非阴影和阴影标准差更低,精度更高;同时,本文算法平均时间花销分别为0.355 s,相比已有文献的结果,本文阴影检测时间花销更少。【结论】提出的基于半监督离散度聚类算法提升了土壤彩色图像阴影检测效率,算法有效。

[Purposes]Due to the influence of light and natural fracture concavity of soil core soil,there are shadows in the soil image acquired by machine vision acquisition terminal.In order to avoid the interference to the subsequent soil species identification,it is necessary to detect the shadow of soil image.[Methods]Through the analysis of shadow and non-shadow in HSI color space of soil color image,it was found that shadow and non-shadow had certain separation characteristics in H and I components respectively.Firstly,in order to increase the separation characteristic for soil image shadow detection,H is converted to H″,and then H″and Iare used to introduce stretching factor to construct m-measure with significant bimodal and separation characteristic for shadow detection.Then,the shadow detection threshold is roughly estimated by using the bimodal characteristic of m-measure histogram.According to the roughly estimated detection threshold and two main peak points,some data are calibrated for shadow and non-shadow,and the supervision information of shadow and non shadow area is obtained.Finally,the data subset to be detected is constructed,and the discreteness of the data subset and the supervision information of shadow and non-shadow is defined.Semi-supervised clustering of uncalibrated data is carried out step by step according to the dispersion.Semi-supervised clustering is used to complete the shadow detection of soil color image.[Findings]The experimental results show that the average standard deviations of non-shadow and shadow in soil image segmentation by comparison algorithms of reference[26-28]are 0.009,0.163;0.182,0.029;0.103,0.087,respectively.There are order of magnitude differences between the standard deviations of non-shadow and shadow of soil image,and the shadow detection of soil color image is failed.The standard deviations of non-shadow and shadow detection algorithms in reference[4]and here are 0.097,0.085;0.063,0.058,respectively.The results show that the both algorithms are effective,but the standard deviation of non-shadow and shadow of soil image detected here is lower,so the accuracy of shadow detection algorithm here is higher.At the same time,the average time cost of the algorithms in reference[4]and here are 0.480 s and 0.355 s,so the time cost of shadow detection here is lower.[Conclusions]It improves the efficiency of shadow detection of soil color image based on semi-supervised discrete degree clustering algorithm,and the algorithm is effective.