为实现温室环境机器人采摘作业中果实的品质判别和空间定位， 研究了基于近红外图像的黄瓜果实识别及特征获取方法。 分析比较黄瓜果实、 茎、 叶在各光谱波段的分光反射特性， 确定采用850 nm干涉滤光片来获取图像， 解决近色系目标、 背景的区分问题； 利用果实的灰度特征， 将P参数阈值法用于图像分割， 实现目标的初步识别， 并对目标图像进行等间距区域化处理， 依据区域块重心、 面积差异滤除噪声、 标记果实； 根据黄瓜的形状纹理特征， 计算果实中心线长度和弯曲度作为黄瓜的质量判别标准， 利用果实与果柄交界处灰度的变化标记出可抓取区域。 通过对温室场景下随机拍摄的包含30幅黄瓜果实图像和10幅无果实图像分别进行算法验证， 结果表明识别准确率各为83.3%和100%， 对抓取区域提取的成功率为83.3%。

For the cucumber harvesting robot, the identification of target information is one of important tasks in the automation of fruit-picking. In order to implement spatial fruit localization and quality discrimination in greenhouse, this paper presented a machine vision algorithm for the recognition and detection of cucumber fruits based on near-infrared spectral imaging. By comparing the spectral reflectance of cucumber plant (fruit, leaf and stem) from visible to infrared region (325-1 075 nm) measured by ASD FieldSpec Pro VNIR spectrometer, a monospectral near-infrared image at the 850 nm sensitive wavelength was captured to cope with the similar-color segmentation problem in complex environment. Then, a method of fruit extraction was developed on the basis of the following steps. Firstly, from the gray level histogram it was observed that the pixels of fruit distributed on the right are lesser than that of background, so “P parameter threshold method” was used to image segmentation. Subsequently, divided local image was partitioned into several sub-blocks by the application of adaptive template mining, which was feasible for processing the fruit with long-column feature. Finally, noises including parts of stem and leaf were eliminated using estimation condition of barycentre position and area size, proved by relative experiment. In addition, the region for robotic grasping was established by gray variation between fruit-handle and fruit pedicel, as the quality feature was extracted with morphological characteristics of the centre-line length and the fruit flexure degree. A detecting experiment was carried out on 30 images with cucumber fruits and 10 images with no fruits, which were taken in a changing greenhouse environment. The results indicate that the accuracy rate of the recognition was 83.3% and 100%, while the success rate of effectively acquiring the grasping region was 83.3%, which can meet the demand of robotic fruit-harvesting.