In order to realize the recognition and localization of the mini-watermelon with stereoscopic cultivation in greenhouse, a machine vision method for acquiring harvesting information of watermelon based on the near-infrared spectral image was presented. By comparing the spectral reflectance of fruit, leaf and stem, a wavelength of about 850nm was chosen as the best wavelength, of which the images taken at different illumination conditions were tested for fruit recognition. At first, the Otsu threshold algorithm was adopted to eliminate most background information. Then, a template liked circle was used to detect fruit region and reduce the noises. Thirdly, according to the morphological feature, the centroid of fruit was considered as the optimum point for picking and the cutting point was judged by “block-location method”. 50 images including fruits and 20 images without fruits were tested by the recognition algorithm, which can satisfactorily detect fruit with a recognition rate of 86% and 95%, respectively, and the accuracy rate of locating algorithm for picking point and cutting point detection was 93.0% and 88.4%, respectively, which met the demand of robotic vision system.