基于对数相似度约束Otsu的自然场景病害果实图像分割

Image Segmentation of Fruits with Diseases in Natural Scenes Based on Logarithmic Similarity Constraint Otsu

针对自然场景下,由于复杂背景以及多变环境,水果病害果实图像分割难的问题,提出了一种基于对数相似度约束Otsu和水平集活动轮廓的近椭圆形病害果实图像分割方法。考虑背景的复杂多变,提出对数相似度约束Otsu分割来区分病害果实与背景;由于水平集活动轮廓模型的局部最优性,提出采用自适应膨胀系数的改进距离规则水平集活动轮廓模型来精确演化轮廓。先对病害果实区域样本的颜色进行混合高斯建模,获得整个病害果实图像与样本模型的对数相似度;对对数相似度进行约束Otsu阈值分割以及形态学滤波;采用最小二乘法对滤波后的曲线轮廓进行椭圆拟合,对拟合后的椭圆采用自适应膨胀系数的距离规则水平集活动轮廓演化,得到病害果实完整轮廓。对18个不同场景的病害果实进行分割,平均误判率和漏判率分别为1.77%和1.6%,实验结果表明,该方法可以从复杂自然场景图像中分割出病害果实。

Due to complex and changeful environment,the image segmentation of fruits with diseases in natural scenes is a difficult problem. A logarithmic similarity constraint Otsu and level set active contour（ LSAC） based image segmentation approach of fruits with diseases was proposed in this paper.Considering the complexity and changeableness in natural scenes, the constraint Otsu method for segmenting logarithmic similarity image between diseased fruits and samples was introduced to distinguish diseased fruits and background; because of the local optimality of LSAC,improved distance regularization level set evolution（ DRLSE） with adaptive expansion coefficient was used to lead contour to actual position. Firstly,the sample color of fruits with diseases,which included not only health area but also diseases area,was modeled using Gaussian mixture model（ GMM）,and then the logarithmic similarity between the image of fruits with diseases and model was obtained. Secondly,logarithmic similarity image was segmented with constraint Otsu and then morphology operator was used to filter out noise and interference. Thirdly,least-squares ellipse fitting method was employed to further removal interference and get initial contour for LSAC. Finally,the contour of fruits with diseases was evolved to the actual position taking use of improved DRLSE with adaptive expansion coefficient. The experimental results show that the actual contour of fruits with disease in complex natural scenes can be obtained and the proposed method can provide the basis for the subsequent diseases density estimation and prevention of fruit diseases.