基于RGBI色彩空间的香蕉图像分割研究

Study on Banana Image Segmentation Based on RGBI Color Space

摘要: 香蕉是我国华南地区主要经济作物，采后品质分级与机器人采摘均需快速、准确地分割果实。传统RGB空间易受光照、阴影及叶片遮挡干扰，分割精度受限。为此，本文提出基于RGBI四维色彩空间的香蕉图像分割方法：在设备原生RGB基础上引入亮度分量I = (R + G + B)/3，实现颜色–亮度联合表征；通过自适应阈值与形态学后处理，在保持实时性的同时显著提升抗光照鲁棒性。试验以480幅温室及田间自然光图像为对象，对比RGB-Otsu、HSV-Otsu、Lab-Kmeans等方法，RGBI-Otsu的Dice系数达0.915，mIoU达0.878，单幅处理时间19.7 ms，优于对照组。进一步将RGBI特征嵌入轻量化UNet，在Jetson Nano边缘端可达18 FPS，满足采摘机器人实时需求。研究表明，RGBI空间在不增加硬件成本的前提下，为香蕉及其他果蔬的精准分割提供了经济高效的新途径。

Abstract: Banana is the main cash crop in South China. After harvesting quality classification and robot picking, the fruit needs to be quickly and accurately divided. The traditional RGB space is easily disturbed by light, shadow and blade occlusion, and the segmentation accuracy is limited. To this end, this paper proposes a banana image segmentation method based on RGBI four-dimensional color space: the brightness component I = (R + G + B)/3 is introduced on the basis of the device’s native RGB to realize the joint color-brightness characterization; through adaptive threshold and morphological post-processing, the anti-light robustness is significantly improved while maintaining real-time. The test took 480 greenhouse and field natural light images as the object. Compared with RGB-Otsu, HSV-Otsu, Lab-Kmeans and other methods, the Dice coefficient of RGBI-Otsu reached 0.915, mIoU reached 0.878, and the single processing time was 19.7 ms, which was better than the control group. The RGBI feature is further embedded in the lightweight UNet, which can reach 18 FPS at the edge of Jetson Nano to meet the real-time needs of picking robots. Research shows that RGBI space provides a new economical and efficient way for the accurate division of bananas and other fruits and vegetables without increasing hardware costs.