Visual Granulometry: Image-based Granulometry of Concrete Aggregate

# Introduction
Concrete is one if the most used building materials worldwide. With up to 80% of volume, a large constituent of concrete consists of fine and coarse aggregate particles (normally, sizes of 0.1mm to 32 mm) which are dispersed in a cement paste matrix. The size distribution of the aggregates (i.e. the grading curve) substantially affects the properties and quality characteristics of concrete, such as e.g. its workability at the fresh state and the
mechanical properties at the hardened state. In practice, usually the size distribution of small samples of the aggregate is determined by manual mechanical sieving and is considered as representative for a large amount of aggregate. However, the size distribution of the actual aggregate used for individual production batches of concrete varies, especially when e.g. recycled material is used as aggregate. As a consequence, the unknown variations of the particle size distribution have a negative effect on the robustness and the quality of the final concrete produced from the raw material.

Towards the goal of deriving precise knowledge about the actual particle size distribution of the aggregate, thus eliminating the unknown variations in the material’s properties, we propose a data set for the image based prediction of the size distribution of concrete aggregates. Incorporating such an approach into the production chain of concrete enables to react on detected variations in the size distribution of the aggregate in real-time by adapting
the composition, i.e. the mixture design of the concrete accordingly, so that the desired concrete properties are reached.


# Classification data
In the classification data, nine different grading curves are distinguished. In this context, the normative regulations of DIN 1045 are considered. The nine grading curves differ in their maximum particle size (8, 16, or 32 mm) and in the distribution of the particle size fractions allowing a categorisation of the curves to coarse-grained (A), medium-grained (B) and fine-grained (C) curves, respectively. A quantitative description of the grain size distribution of the nine curves distinguished is shown in the following figure, where the left side shows a histogram of the particle size fractions
0-2, 2-8, 8-16, and 16-32 mm and the right side shows the cumulative histograms of the grading curves (the vertical axes represent the mass-percentages of the material).

For each of the grading curves, two samples (S1 and S2) of aggregate particles were created. Each sample consists of a total mass of 5 kg of aggregate material and is carefully designed according to the grain size distribution shwon in the figure by sieving the raw material in order to separate the different grain size fractions first, and subsequently, by composing the samples according to the dedicated mass-percentages of the size distributions.


For data acquisition, a static setup was used for which the samples are placed in a measurement vessel equipped with a set of calibrated reference markers whose object coordinates are known and which are assembled in a way that they form a common plane with the surface of the aggregate sample. We acquired the data by taking images of the aggregate samples (and the reference markers) which are filled in the the measurement vessel and whose constellation within the vessel is perturbed between the acquisition of each image in order to obtain variations in the sample’s visual appearance. This acquisition strategy allows to record multiple different images for the individual grading curves by reusing the same sample, consequently reducing the labour-intensive part of material sieving
and sample generation. In this way, we acquired a data set of **900 images** in total, consisting of 50 images of each of the two samples (S1 and S2) which were created for each of the nine grading curve definitions, respectively (50 x 2 x 9 = 900). For each image, we automatically detect the reference markers, thus receiving the image coordinates of each marker in addition to its known object coordinates. We make use of these correspondences for the computation of the homography which describes the perspective transformation of the reference marker’s plane in object space (which corresponds to the surface plane of the aggregate sample) to the image plane. Using the computed homography, we transform the image in order to obtain an perspectively rectified representation of the aggregate sample with a known, and especially a for the entire image consistent, ground sampling distance (GSD) of **8 px/mm**. In the following figure, example images of our data set showing aggregate samples of each of the distinguished grading curve classes are depicted.


## Related publications:
If you make use of the proposed data, please cite the publication listed below.

* **Coenen, M., Beyer, D., Heipke, C. and Haist, M., 2022**: Learning to Sieve: Prediction of Grading Curves from Images of Concrete Aggregate. In: _ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences V-2-2022_, pp. 227-235, [Link](https://doi.org/10.5194/isprs-annals-V-2-2022-227-2022).

# 引言
混凝土是全球应用最广泛的建筑材料之一。其体积占比可达80%的主要组分为粗细骨料颗粒（通常粒径范围为0.1 mm至32 mm），这些颗粒分散于水泥浆基体中。骨料的粒径分布（即级配曲线(grading curve)）对混凝土的性能与质量特征具有显著影响，例如新拌状态下的和易性以及硬化状态下的力学性能。实际生产中，通常通过人工机械筛分测定小批量骨料的粒径分布，并将其作为大批量骨料的代表性参数。然而，混凝土各生产批次实际使用的骨料粒径分布存在波动，尤其当采用再生骨料作为原材料时波动更为明显。由此产生的粒径分布未知波动会对最终成品混凝土的生产稳定性与质量产生负面影响。

为精准掌握骨料实际粒径分布，从而消除材料性能的未知波动，我们提出了用于基于图像预测混凝土骨料粒径分布的数据集。将该方法融入混凝土生产链后，可实时响应检测到的骨料粒径分布变化，同步调整混凝土配合比设计，进而达成预期的混凝土性能目标。


# 分类数据集
本分类数据集共区分9种不同的级配曲线(grading curve)，所有曲线均符合DIN 1045的规范要求。这9种级配曲线的差异体现在最大粒径（8 mm、16 mm或32 mm）以及粒级分布上，据此可将曲线分为粗粒级（A类）、中粒级（B类）与细粒级（C类）三类。下文将通过图表定量描述这9种级配曲线的粒径分布：左侧为粒级0-2 mm、2-8 mm、8-16 mm及16-32 mm的直方图，右侧为级配曲线的累积直方图（纵轴代表材料的质量百分比）。

针对每种级配曲线，我们制备了两份骨料样品（S1与S2）。每份样品总质量为5 kg，制备流程为先通过筛分分离原材料的不同粒级，再按照图示粒径分布的指定质量百分比进行配比，从而得到精心设计的目标样品。


数据采集环节采用静态采集装置：将样品置于测量容器中，容器内安装有一组校准参考标记，其物方坐标已知，且参考标记与骨料样品表面共面。我们通过拍摄装填入测量容器的骨料样品（含参考标记）的图像完成数据采集，每次拍摄前调整样品在容器内的摆放位置，以获取样品视觉外观的多样性变化。该采集策略可复用同一样本为对应级配曲线拍摄多张不同图像，从而大幅减少筛分与制样的繁重工作量。最终我们共采集得到**900张图像**：针对9种级配曲线，每种曲线对应的两份样品（S1与S2）各拍摄50张图像，总计50×2×9=900张。

针对每张图像，我们自动检测参考标记，可同时获取每个标记的图像坐标与已知物方坐标。基于这些坐标对应关系，我们计算得到单应性矩阵(homography)，该矩阵可将参考标记所在的物方平面（即骨料样品表面平面）的透视变换关系映射至图像平面。通过该单应性矩阵对图像进行校正，可得到透视校正后的骨料样品图像，其地面采样距离(Ground Sampling Distance, GSD)统一为**8像素/毫米（8 px/mm）**。下图展示了本数据集中各等级配曲线类别的骨料样品示例图像。


## 相关引用文献
若您使用本数据集，请引用以下文献：
* **Coenen, M., Beyer, D., Heipke, C. 和 Haist, M., 2022**：《Learning to Sieve: Prediction of Grading Curves from Images of Concrete Aggregate》（《学习筛分：基于混凝土骨料图像的级配曲线预测》）。发表于《ISPRS摄影测量、遥感与空间信息科学年鉴 V-2-2022》，页码227-235，[链接](https://doi.org/10.5194/isprs-annals-V-2-2022-227-2022)。