基于深度学习的细骨料图像实时分割提取

文章基于深度学习算法对细骨料投影图像进行分割, 通过对比传统阈值分割与 PSPNet、DeepLab V3+、U-Net 深度学习网络模型算法的分割结果对 4 种模型进行评价分析, 同时实验对比细骨料 2 种等效粒径计算方法(单面投影法、双面投影法)的粒径和级配分布结果。结果表明: 深度学习模型算法中 U-Net 网络模型的准确率、召回率、F 平衡分数和交并比分别达到 99.8%、88.1%、84.9%、84.3%, 均优于对比组模型; 对于 3 种不同粒径的单粒段细骨料, 采用双面投影法计算出的细骨料等效粒径 $ D_{d} $ 与实际细骨料粒径的偏差分别为 1.40%、2.10%、3.12%; 对于混合粒段骨料, 采用等效粒径 $ D_{d} $ 计算出的级配分布曲线更接近筛分法的实验结果, 具有普遍适用性。研究结果可为细骨料径粒径和粒型参数的计算提取提供新的思路。

In this paper, the deep learning algorithm is used to segment the fine aggregate projection image, and the evaluation and analysis on the traditional threshold segmentation and three deep learning network model algorithms (PSPNet, DeepLab V3+ and U-Net) are conducted by comparing their segmentation results. At the same time, the results of grain size and gradation distribution of fine aggregate measured by two equivalent grain size calculation methods (single-sided projection method and double-sided projection method) were compared experimentally. The results show that the accuracy rate, recall rate, F-balance score and intersection ratio of U-Net network model in the deep learning model algorithm are 99.8%, 88.1%, 84.9% and 84.3%, respectively, which are superior to those of the control group model. For the single-grain segment fine aggregate with three different grain sizes, the deviation between the equivalent grain size $ D_{d} $ of fine aggregate calculated by double-sided projection method and the actual fine aggregate size is 1.40%, 2.10% and 3.12%, respectively. For the aggregate of mixed grain segment, the gradation distribution curve calculated by $ D_{d} $ is closer to the experimental results of screening method, which has universal applicability. The results provide a new idea for the study of grain size and grain type parameters of fine aggregate.