基于深度学习的多品种鲜食葡萄采摘点定位

鲜食葡萄品种多样，具有不同的形状和颜色。针对葡萄采摘机器人采摘不同品种鲜食葡萄时采摘点定位精度降低的问题，提出一种基于深度学习的多品种鲜食葡萄采摘方法。首先利用PSPNet(MobileNetv2)语义分割模型分割葡萄图像，在葡萄上方设置一个兴趣区域，在兴趣区域内使用自适应阈值果梗方向Canny边缘检测提取果梗边缘信息，然后采用霍夫变换检测果梗边缘上的直线段并进行直线拟合。最后将拟合的直线与兴趣区域的水平对称轴的交点作为采摘点。对晴天顺光、晴天逆光、晴天遮阴3种光照条件下的克瑞森、阳光玫瑰、红提和黑金手指4个品种的360幅葡萄图像进行采摘点定位试验。结果显示，采摘点定位准确率为91.94%,定位时间为187.47 ms,在模拟试验中采摘成功率为85.5%。

Picking point positioning of multi variety table grapes based on deep learning

Fresh grapes come in a variety of different shapes and colors. To address the problem of reduced accuracy of picking point positioning when the grape picking robot picks different varieties of fresh grapes, a multi variety fresh grape picking method based on deep learning is proposed. In this paper, firstly, a PSPNet (MobileNetv2) semantic segmentation model is used to segment the grape image, a region of interest is set above the grapes, the stalk edge information is extracted within the region of interest using adaptive thresholding of the stalk direction Canny edge detection, then the straight line segment on the stalk edge is detected using the Hough transform and a straight line is fitted. Finally the intersection of the fitted straight line with the horizontal symmetry axis of the region of interest is taken as the picking point. The picking point location experiment was carried out on 360 grape images of four varieties of Crescent, Sun Rose, Red Raisin and Black Goldfinger under three lighting conditions as sunny day with light, sunny day with light backlight and sunny day with shade. The results showed that the picking point positioning accuracy was 91.94%, the positioning time was 187.47 ms and the picking success rate was 85.5% in the simulated experiments. This study achieves rapid picking point positioning for multiple varieties of fresh grapes and provides technical support for picking point positioning in grape picking robots.