基于RGBD相机的黑皮鸡枞菌子实体形态视觉测量

子实体的形态是黑皮鸡枞菌育种和栽培过程中关注的重点，传统手工测量方法耗时、费力。为了实现在生长过程中对黑皮鸡枞菌子实体进行快速、准确地监测，提出了一种基于SR300深度相机的自动化在线测量方法。首先，在实验室和栽培工厂采集样本、构建数据集并基于YOLO V4建立黑皮鸡枞子实体识别模型。其次，根据K-means算法对识别所得的包围框分布情况聚类，并以此进行分割，得到最优处理区域后上采样构建不同尺度的子实体图像；最后，融合Grabcut算法提取的不同尺度目标轮廓边界得到测量点信息，并结合深度值矫正待测点的位置，计算真实值。经过现场试验，结果表明该方法能够得到子实体的形态特征，菌盖测量平均绝对误差为0.357 mm，标准差为0.273 mm；菌柄宽度测量平均绝对误差为0.304 mm，标准差为0.209 mm。该方法能够满足工厂中对黑皮鸡枞菌子实体形态自动化监测的需求。

Measurement of morphology of oudemansiella raphanipies based on RGBD camera

Abstract: Oudemansiella raphanipies is one of the most popular mushroom with the various vitamins, dietary fiber, and a large amount of protein for the body healthy and immunity. The growth state of Oudemansiella raphanipies depends mainly on the morphology of the sporophore during the cultivation and harvesting. The greenhouse environment can also be optimized for the noticing changes of the sporophore in the best way. However, the manual measurement of the sporophore morphology cannot fully meet the large-scale production in recent years, due to the high cost and labor intensity. Furthermore, the measuring standards are also varied greatly. It is a high demand for the advanced information technology. In this study, an automatic online measurement was proposed to rapidly and accurately monitor the sporophore of Oudemansiella raphanipies during the growth using SR300 depth camera. Firstly, the samples were collected both in the greenhouse of Bioengineening and Technological Research Centre for the Edible and Medicinal Fungi in the Jiangxi Agricultural University and the cultivation factories located in Zhangshu and Nanchang City, Jiangxi, China. The data sets were obtained to divide into 7 (training set): 3 (test set) after the image cleaning and enhancement, as well as labelling tags by LabelMe software. An identification model was established to accurately locate the sporophore using the YOLO V4 target detection. Secondly, the distribution of the bounding boxes was clustered, according to the K-means algorithm, and thereby the more optimized processing area was gained by the segmentation. Afterwards, the images of different scales were generated by the up-sampling operation. The contour boundary curves were extracted to integrate the different scales with the Grabcut algorithm. Meanwhile, the true value was calculated, after the sporophore contour was corrected with the depth information. Experiments showed that a faster and better way was realized to extract the morphological characteristics of the cap and stipe. The average accuracy of the YOLO V4 model in the test set was 91.17%, with the recall rate 84.08%, and the F1 score 0.87. According to the quantity of the sporophore, the average time was promoted by 95.37%, 95.17%, and 81.92%, respectively, on the segmentation of optimal processing area, compared with the original images, Besides, the rates of error and repeatability of the measurement were promoted dramatically, as the distance from the target increased. Therefore, the distance between the target and SR300 should be no more than 40cm. Compared with the manual, the best performance was achieved, with the coefficient of determination of 0.979 and 0.977 for the cap and stipe, respectively, the average error of the cap and the stipe''s width 0.357, and 0.304mm, 7, respectively, and the STD of 0.273 and 0.209 mm, respectively. Therefore, this measurement can fully meet the high requirements of the low-cost and high-precision automatic monitoring on the morphology of Oudemansiella raphanipies at the edible mushroom cultivation site. The findings can also lay the foundation for the real-time and accurate measurement of Oudemansiella raphanipies in a complex environment using the picking robots.