机构地区:[1]福建工程学院计算机科学与数学学院,福州350118 [2]福建工程学院交通运输学院,福州350118 [3]福建金森林业股份有限公司,三明353300 [4]福建工程学院机械与汽车工程学院,福州350118
出 处:《林业工程学报》2022年第2期135-142,共8页Journal of Forestry Engineering
基 金:福建省科技厅自然科学基金(2018J01619);福建金森林业股份有限公司校企合作项目(GY-H-20154);福建省林业科技项目(2021FKJ06)。
摘 要:针对木材检尺中采用人工检尺的方法存在效率低下且检尺主观性较强的问题,提出一种基于掩模区域卷积神经网络(Mask R-CNN)实例分割模型的木材分割方法,探究实例分割在木材密集堆放场景下对各尺寸木材分割的可行性,以期实现智能检尺,提高检尺效率。应对密集木材检测分割任务,难点在于密集小木材和大木材的检测。本研究在原始Mask R-CNN模型的基础上通过改进优化模型参数,包括多尺度训练、提升样本采样数、提高图片输入尺寸和有效数据增广等技术,进行多组分割对照实验,同时利用OpenCV库对模型输出的木材分割掩码图完成木材轮廓拟合和木材计数,并就实验结果对模型性能进行分割精度、掩码质量和木材真检率等多维度分析。实验结果表明:测试集木材真检率达到97.989%,误检率为0.30%,并且相较基础网络,对小木材和大木材的检测分割能力提升明显,分割精度最佳提升12.9%和5.2%,掩码分割质量最佳提升2.2%。改进后的Mask R-CNN模型对密集场景下的木材分割效果良好,此外算法具有较强的鲁棒性及迁移能力,微调下能适应各种场景下的各尺寸大小密集木材检测分割任务。Aiming at solving the problems of low efficiency and strong subjectivity in the manual measurement method of measuring diameters of log end faces,a segmentation method of dense-stacked logs using the Mask Region-based Convolutional Neural Network(Mask R-CNN) instance segmentation model is proposed to explore how the instance segmentation model can be used in scenes of dense-stacked logs.The feasibility of dividing stacked logs of various sizes is expected to realize the intelligent measurement of log diameters,improve the efficiency of log diameter measurements and reduce the cost of measurement.In this dense-stacked logs detection and segmentation task,the difficulty lies in the detection of dense-stacked small logs and large logs.Due to the poor performance of the original Mask R-CNN model in detection and segmentation of dense-stacked small log and large log targets,this study optimizes the model parameters on the basis of the original Mask R-CNN model in four aspects,including multi-scale training of the input image size model,increasing the sample number of the model Region proposal network and Region-based Convolutional Neural Network modules,increases the input size of the model training image,and performs effective data augmentation on the training images.In view of these four optimization methods and other log detection reference methods,a multi group detection and segmentation control experiment was carried out.At the same time,the OpenCV-Python third-party library was used to complete the log contour search,log contour fitting and log counting on the log segmentation mask map output by the model.Finally,based on the experimental results,the multi-dimensional analysis of the model performance was carried out,such as the segmentation accuracy of model training,the quality of model segmentation mask and the true detection rate of model logs.The experimental results show that the model’s true detection rate of wood on the test set images reaches 97.989%,and the false detection rate is 0.30%.Compared with
关 键 词:密集木材检测 木材分割 Mask R-CNN 木材计数 深度学习
分 类 号:S781[农业科学—木材科学与技术] TP391.41[农业科学—林学]
正在载入数据...
正在载入数据...
正在载入数据...
正在载入数据...
正在载入数据...
正在载入数据...
正在载入数据...