【作者】 罗泽；

【导师】 孙丽萍； 张怡卓；

【作者基本信息】 东北林业大学 ， 林业工程自动化， 2022， 博士

【摘要】 红松是中国东北地区最重要的经济树种,松果采摘是红松林管理中的一个重要问题,目前这项工作主要依靠人工完成。人工采摘不仅要求工人具备较高的专业技能,同时还存在成本高、风险大的缺点。研发一种有效的松果检测方法,对松果进行探测和定位,并将其集成于采摘机器人,可以为松果的机械化收获提供技术支持。然而,由于目标对象小和检测环境复杂等因素的影响,松果检测存在一定的困难,学术界的相关研究很少。现有的松果检测方法也是基于传统方法,存在检测速度慢和检测精度低的缺点。随着计算机视觉的发展,使用深度学习技术解决目标检测问题已成为当前研究的主流方向,其检测速度和检测精度都大幅度超越了传统的检测方法。本文在分析国内外研究现状的基础上,围绕检测模型训练数据不足、小目标松果对象检测精度不高和算法检测速度欠佳等问题,研究了一种基于深度学习的小目标松果检测方法,并将其与所设计的松果采摘机器人进行了系统集成,具体研究成果与创新点包括:提出了一种基于生成对抗网络的小目标松果图像数据增强模型(PC-GAN),解决了松果图像数据采集困难和检测模型训练数据量不足的问题。该模型通过调整生成图像的尺寸和潜在噪声向量的维数,使得生成图像的数据分布具有连续性,从而增强生成图像区域特征的真实性;模型网络设计中利用自动编码器结构提高生成图像的纹理清晰度,使用长跳跃连接增强网络间的浅层信息流动,从而实现生成图像中松果的细节特征补全;设计基于Wasserstein距离的损失函数,以解决生成对抗网络中的模式崩塌问题和训练不稳定问题,并结合空间约束、结构相似性算法和鉴别器的重构损失进一步提升模型的图像生成性能。提出了一种基于多尺度特征和模型压缩技术改进的YOLOv4小目标松果检测算法,提高了小目标松果对象的检测精度和检测速度。该算法针对小目标松果的特征提取问题,利用密集连接和跨阶段局部网络增强小目标语义信息在网络中的传递,并通过分层结构实现多尺度特征提取;针对小目标松果的特征融合问题,设计基于尺度均衡金字塔原理的多尺度特征融合方法,通过3D融合实现不同尺寸特征图之间的相互利用,从而增强特征融合的效果;针对算法检测速度欠佳的问题,设计基于通道剪枝的模型压缩方法,对不同位置的跳跃连接使用不同的修剪方式,通过多次迭代的剪枝方案实现模型压缩率和检测精度的平衡,并结合批量归一化折叠技术进一步提升模型的压缩效果。构建了一种基于采摘机器人机器视觉的松果检测系统,实现了小目标松果检测算法与机械采摘设备的系统集成。首先,通过分析松果采摘任务的技术需求确定移动平台、视觉模块和采摘机械臂的选择方案。然后,针对红松林中的实际作业环境设计采摘机械臂和末端执行器的机械结构,搭建松果采摘机器人原型机的机械模型。最后,针对松果采摘机器人的任务需求设计相应的松果检测系统,将松果检测算法与松果采摘机器人进行系统集成。试验结果表明:PC-GAN生成的小目标松果图像IS分数和FID分数分别为7.78和19.01,具备很好的图像质量;生成图像经过图像融合后可以有效实现松果图像数据集的扩充,从而提升检测算法的精确性和鲁棒性;改进后YOLOv4检测算法的AP值提高了6%,模型的计算成本和存储需求分别降低了48.2%和38.8%,可以显著提高小目标松果对象的检测精度和检测速度;将研发的小目标松果检测算法与松果采摘机器人进行系统集成后,可以对松果目标实现有效的检测,完成模拟环境下的松果采摘任务。本文的研究成果可以为松果采摘设备的研制提供技术支持,具有借鉴意义。更多还原

【Abstract】 Korean pine is the most important economic tree species in Northeast China.Pinecone picking is an important issue in the management of Korean pine forests.At present,this work is mainly done manually.Manual picking not only requires workers to have high professional skills,but also has the disadvantages of high cost and high risk.To develop an effective detection method to detect and locate the pinecone,and integrate it into the picking robot,which can provide technical support for the mechanized harvesting of pinecone.However,due to the influence of factors such as small target and complex detection environment,there are some difficulties in pinecone detection,and few relevant studies in academic circles.The existing methods of pinecone detection are also based on traditional methods,which have the disadvantages of slow detection speed and low detection accuracy.With the development of computer vision,the use of deep learning technology to solve the problem of object detection has become the mainstream direction of current research,and its detection speed and detection accuracy are greatly beyond the traditional detection methods.Based on the analysis of the research status at home and abroad,this paper focusing on the problems insufficient training data of detection model,low detection accuracy of small target pinecone object,and poor detection speed of algorithm,studies a small target pinecone detection method based on deep learning,and systematically integrates it with the designed pinecone picking robot.Specific research achievements and innovations include:A small target pinecone image data enhancement model based on generation adversarial network(PC-GAN)is proposed to solve the problem of difficult pinecone image data acquisition and insufficient training data set of detection model.In this model,the size of the generated image and the dimension of the potential noise vector are adjusted to make the data distribution of the generated image continuous,so as to enhance the authenticity of image region features.In the model network design,the automatic encoder structure is used to improve the texture definition of the generated image,and the long jump connection is used to enhance the shallow information flow between networks,so that the small target pinecone details are completed.The loss function based on Wasserstein distance is designed to solve the problem of model collapse and training instability in the generation adversarial network.Furthermore,the spatial constraint,structural similarity algorithm and discriminator reconstruction loss are combined to further improve the image data generation performance of the model.This paper proposes an improved YOLOv4 small target pinecone detection algorithm based on multi-scale features and model compression technology,which improves the detection accuracy and detection speed for small target pinecone objects.Aiming at the feature extraction problem of small target pinecones,use dense connection and cross-stage local network to enhance the transfer of semantic information of small target pinecones in the network,and then multi-scale feature extraction was achieved through hierarchical structure.Aiming at the problem of small target pinecone feature fusion,a multi-scale feature fusion method based on scale equilibrium pyramid principle was designed.The 3D fusion was used to realize the mutual utilization of different size feature maps,so as to enhance the effect of feature fusion.Aiming at the problem of poor algorithm detection speed,a model compression method based on channel pruning was designed.This method uses different pruning methods for jumping connections at different positions,achieves the balance between compression rate and detection accuracy through multiple iterations of pruning scheme,and further improves the compression effect of the model by combining with batch normalized folding technology.A pinecone detection system based on the machine vision of the picking robot was constructed,and the integration of the pinecone detection algorithm with the picking machinery was realized.Firstly,the technical requirements of pinecone picking task were analyzed to determine the selection scheme of mobile platform,vision module and picking robot arm.Secondly,the mechanical structure of the picking robot arm and end-effector was designed according to the actual working environment in the Korean pine forest,and the physical model of the prototype pinecone picking robot was constructed.Finally,the corresponding pinecone detection system was designed according to the task function requirements of the pinecone picking robot,and the pinecone detection algorithm was integrated with the picking robot.The experimental results show that the IS score and FID score of small target pinecone images generated by PC-GAN are 7.78 and 19.01,respectively,which have good image quality.After image fusion,the generated image can effectively expand the pinecone image data set,thus improve the accuracy and robustness of the detection algorithm.The AP value of the improved YOLOv4 detection algorithm is increased by 6%,the calculation cost and storage demand of the model are reduced by 48.2% and 38.8%,respectively,which significantly improves the detection accuracy and detection speed of small target pinecone objects.The pinecone detection algorithm of small target is integrated with the pinecone picking robot,which can effectively detect the pinecone target,and complete the pinecone picking task under the simulated environment.The research results of this paper provide technical support for the development of pinecone picking equipment and has reference significance.更多还原