自然环境下重叠与遮挡苹果图像识别方法研究  

Image recognition algorithm research of overlapped apple fruits in the natural environment

作　　者：刘丽娟[1] 窦佩佩 王慧[1] Liu Lijuan;Dou Peipei;Wang Hui(College of Information Science and Technology,Shihezi University,Shihezi,832003,China)

机构地区：[1]石河子大学信息科学与技术学院,新疆石河子832003

出　　处：《中国农机化学报》2021年第6期174-181,共8页Journal of Chinese Agricultural Mechanization

基　　金：石河子大学校级资助项目(ZZZC201748B)。

摘　　要：针对自然环境下重叠与遮挡苹果的图像识别问题,提出一种组合优化解决方案。该组合算法首先选取被枝叶遮挡或者相互重叠的成熟苹果图像,对图像进行预处理;然后提取lab颜色空间a分量,YUV空间U、V分量进行图像分割,利用改进的多通道全局阈值分割算法获得苹果目标二值图像;最后针对苹果类圆但不规则的性质,采用霍夫变换和圆形约束法识别并定位单一苹果目标。试验结果表明:所提出的识别算法可以在多种自然环境下从重叠与遮挡苹果图像中识别出单个苹果目标;在分割算法方面平均精确度95.5%,平均假阳性率2.1%,运行速度比Kmeans迭代算法快8.9倍,与直接Otsu和Kmeans迭代算法相比精度高、速度快;在圆形识别方面,圆形约束法可在霍夫变换法检测圆后剔除过分割的圆,准确定位目标。A combined algorithm for image recognition and boundary segmentation of overlapped apple fruits under natural environments was proposed.Firstly,the algorithm selects the mature apple images that are shaded by branches and leaves or overlapped with each other and preprocess the images.Then,a component of lab space,U component of YUV space,and V component of YUV space were extracted as image segmentation.The binary image was obtained by using the improved multi-channel global threshold segmentation algorithm.Finally,the Hough transform and Circular constraints are used to identify the individual apple target.The experimental results showed that the proposed algorithm could recognize the overlapped objectives under natural environments.In terms of the segmentation algorithm,the average accuracy was 95.5%,and the average false positive rate was 2.1%.It ran 8.9 times faster than Kmeans iterative algorithm.It had high precision and fast speed comparing with the direct Otsu and Kmeans iterative algorithm.In the aspect of circle recognition,the circle constraint method can eliminate the over-segmented circle after the Hough transform algorithm and accurately locate the target.

关 键 词：果实识别 自然环境 阈值分割 霍夫变换 圆形约束 目标定位 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] S661.1[自动化与计算机技术—计算机科学与技术]