Abstract: In view of the backward automation, low picking efficiency, and short picking cycle in the process of picking Camellia oleifera fruit in China, the machine vision technology applied to robotic harvesting technology is limited by the interference of the complex background in the real scene, which leads to the problem of low recognition accuracy. This paper takes Camellia oleifera fruit in the natural environment as the research object, and proposes an algorithm for identifying and detecting Camellia oleifera fruit in natural scenes based on Mask-RCNN. Firstly, the image of the Camellia oleifera fruit was obtained and the data set was established, and the ResNet convolutional neural network was used to extract the features of the Camellia oleifera fruit image, so as to obtain the fruit target segmentation result, and then RPN was used to operate the obtained feature map, and the full link layer was added to extract the mask pixel area of each sample is, and the target category was predicted. The test set was used to test the segmentation network model and target recognition algorithm of Camellia oleifera fruit respectively. The results showed that the segmentation accuracy of the network model was 89.85%, the average detection accuracy of the target recognition of Camellia oleifera fruit was 8942%, and the recall rate was 92.86%. This algorithm can automatically detect the target of Camellia oleifera fruit, and effectively reduce the interference of factors such as leaf and flower bud occlusion, fruit overlap, fruit color and other factors under different lighting conditions, and provides reliable visual support for automatic fruit picking in natural scene.

Key words: natural scene, Camellia oleifera fruit, target recognition, Mask-RCNN, image segmentation

摘要： 针对我国油茶果采摘过程中存在的自动化水平落后、采摘效率低、适采周期短的现状，应用于机器人收获技术的机器视觉技术受限于真实场景中复杂背景干扰从而导致识别精度较低的问题。以自然场景下的油茶果为研究对象，提出一种基于Mask-RCNN的自然场景下油茶果目标识别与检测算法，首先获取油茶果图像并建立数据集，利用ResNet卷积神经网络提取油茶果果实图片的特征，获得果实目标分割结果，再采用RPN对所得到的特征图进行操作，并增加全连接层，提取每个样本mask像素面积，并对目标类别进行预测。利用测试集分别测试油茶果的分割网络模型及目标识别算法，结果表明，网络模型的分割准确率为89.85％，油茶果目标识别的平均检测精度为89.42％，召回率为92.86％。本算法能够自动检测油茶果目标，并有效降低不同光照情况下叶片与花苞遮挡、果实重叠、果实色泽等因素干扰，为自然场景中果实自动化采摘提供可靠的视觉支持。

关键词: 自然场景, 油茶果, 目标识别, Mask-RCNN, 图像分割