Computer Vision-Based Adhesion Quality Inspection Model for Exterior Insulation and Finishing System

Abstract

The adhesion quality is one of the critical factors affecting EIFS quality. The traditional visual inspection method is time-consuming and labor-intensive work because limited inspectors have to check all insulation units when they are installed. Thus, this study proposes a CV-based automatic adhesion quality-inspection model for EIFS that can quickly and quantitatively inspect the shape and amount of adhesive. In the case study, the proposed models accurately determined the adhesion quality with measurement accuracies of 90.58%, 95.28%, and 93.58% in ribbon width, dab diameter, and adhesive area, respectively, and a root mean-square error of 6.9 mm in dab distance. The results of this study can achieve automation of inspection for EIFS.

1. Introduction

Numerous nations around the globe are actively pursuing strategies to reduce greenhouse gas emissions in a bid to achieve carbon neutrality by 2050. A significant contributor to these efforts is the building sector, which accounts for approximately 40% of global energy consumption [1,2,3]. One promising solution to address challenges in the building sector is the realization of Zero-Energy Buildings (ZEBs), where the Exterior Insulation and Finishing System (EIFS) plays a pivotal role.

EIFS, initially developed in Europe under the name External Thermal Insulation Composite System (ETHICS), is a cladding method that provides an insulated and finished exterior wall surface. Over time, these features have been integrated into EIFS, including waterproofing and drainage functions, refining it into a comprehensive exterior insulation system [4]. This system is particularly effective in reducing heat loss, which accounts for more than half of a building’s total energy demand [3,4,5,6].

The adhesion quality is one of the critical factors affecting the overall performance of EIFS. The most common method of attaching insulation to the structure is by using an adhesive similar to mortar [7]. Poor adhesion quality can result in the detachment of insulation from the wall and pose fire safety risks due to the chimney effect caused by unbonded spaces between the insulation and the wall [8]. Due to these safety concerns, many countries and organizations have established regulations for EIFS adhesion quality. These regulations include specifications regarding adhesive quantity, shapes, bonding strength, cracks, and moisture curing, among other factors.

The traditional inspection method, which involves inspectors checking visually whether the shape and amount of adhesive meet EIFS specifications, makes it difficult to ensure the adhesion quality of EIFS. It is time-consuming and labor-intensive work because limited inspectors have to check all insulation units when they are installed. Also, inspectors can make subjective judgements by visual checks.

Recently, with the development of deep learning, computer vision (CV)-based approaches can aid supervision in various construction tasks, serving as a substitute for human eyes. Many researchers are working on applying deep-learning-based CV techniques to the construction industry. However, applying CV to the adhesion quality of EIFS has not been sufficiently explored.

To address this gap, this study proposes a CV-based adhesion quality-inspection model for EIFS. As previously mentioned, while EIFS adhesion quality includes various factors, this study focuses on enabling CV techniques to quickly and quantitatively inspect the shape and amount of adhesive.

The proposed model consists of three steps: first is the segmentation step to extract the insulation and adhesive from the background. The second is the image transformation step to convert the tilted insulation image into a front-facing image, and the third is the measurement step to measure the shape and amount of adhesive.

The remainder of the paper is organized as follows. Section 2 reviews the relevant studies and EIFS specifications to demonstrate the contributions of the study. Section 3 presents the detailed model proposed in this study. In Section 4, the case study is conducted to validate the proposed model. Finally, Section 5 presents the summary, contributions, and limitations.

2. Literature Review

2.1. CV-Based Quality Management in Construction

In the early stages of CV, conventional image processing techniques and simple classification models based on image classification were initiated. Lin et al. [9] proposed tile alignment inspection methods using conventional image processing techniques such as FAST corner detection. Silva et al. [10] used unmanned aerial vehicles (UAVs) to acquire concrete crack images and applied an image classification algorithm. Wang et al. [11] utilized image classification combined with IoT technology to monitor the vibration quality of fresh concrete.

With the development of deep-learning-based CVs, such as the Convolutional Neural Network (CNN), research on quality management using object detection has also increased. Cha et al. [12] proposed a crack damage detection model using CNN. To assess rebar corrosion, Jeon et al. [13] used Electrical Resistance Tomography (ERT) and detected rebar positions from ERT images.

The advancement of datasets and computer vision, coupled with the development of hardware such as CPU and GPU, which enables more extensive and sophisticated training, has facilitated the use of pixel-level segmentation, which is more advanced compared to early image classification and object detection. Tuan et al. [14] used cascade Mask R-CNN to analyze the particle shape of coarse aggregates. Zeng et al. [15] utilized the PANet model to capture fly ash microspheres. Vasquez et al. [16] performed semantic segmentation for defect detection in window frames for construction-quality inspection. However, these methods of quality inspection have limitations as they mainly focus on location identification or presence detection without performing inspections through quantitative calculations.

Yoon et al. [17] proposed a monitoring model for the application quality of external insulation adhesives using the instance segmentation model YOLOv5-seg, but it has limitations in that it only measures limited items such as the area of the adhesive and the distance between dabs. Therefore, this study aims to propose an automated inspection model for EIFS adhesion quality that can supervise compliance with overall items specified in EIFS regulations, including ribbon thickness and dab diameter.

2.2. EIFS Specifications Related to Adhesion Quality

As mentioned earlier in the introduction, many factors influence adhesion quality. However, this paper specifically focuses on the specifications related to the quantity and shape of adhesive, which can be evaluated using CV. To this end, a literature review was conducted on EIFS specifications concerning the quantity and shape of adhesive as applied in construction sites.

There are several specifications concerning EIFS. In the USA, the American Society for Testing and Materials International (ASTM) covers board sizes through ASTM E 2430 [18] and the minimum requirements for EIFS adhesives through ASTM C 1397-13 [19]. According to ASTM regulations, insulation board size is designated as a length of 48 in. (1219 mm) ± 1/16 in. (1.6 mm) and width of 24 in. ± 1/16 in. (1.6 mm). In the ribbon and dab method, ribbon thickness is 2 in. (50.8 mm), the diameter of the ribbon is 4 in. (101.6 mm), and intervals of the dabs are 8 in. (203.2 mm).

The European Organization for Technical Approvals (EOTA) issued the ETAG 004 [20], which is in Section 6.1.4.1.3. “Bond strength between adhesive and insulation product” specifies the area calculation formulas for the bonding performance of each adhesive (Table 1 and Equation (1)). However, it lacks detailed descriptions regarding the application methods (shapes) of the adhesive.

To comply with the minimal admissible bonded surface requirement described below:

(1)$\mathrm{S}\left(\mathrm{\%}\right) = {\displaystyle \frac{0.03}{B}}\times 100$

where S is the minimal bonded surface, which shall exceed 20%. B is the minimum failure resistance of the adhesive to the insulation product in dry conditions for all failure modes (MPa), and 0.03 MPa is the minimum requirement. According to the table and equation in ETAG-004, in the case of the minimum bonding strength of 0.03 MPa, a 100% bonded surface area is required. Even with an increase in bonding strength, a minimum bonded surface area of 20% is still mandated.

In the Republic of Korea, the Korea Land & Housing Corporation (LH) defined requirements for EIFS by the specification for Housing Projects 46030 in 2012 [21]. In 2020, it was revised to align with the Korea Construction Standard and rebranded as LHCS. Figure 1a shows the previous version of the specification. LH46030 stated that insulation boards should not exceed 600 $\times$ 1200 mm, the adhesive should cover at least 40% of the surface, and the specified ribbon width should comprise 50 mm and 100 mm diameter of six dabs, with distances between each dab.

In the revised specification LHCS 41 42 00 15 [22] Figure 1b, following the LH specification for Housing Projects 46030, the insulation board size of 600 × 1200 mm, the adhesive area, the ribbon thickness of 50 mm, and the dab diameter of 100 mm were maintained. On the other hand, the number of dabs increased to eight, while the distance between each dab was removed.

Table 2 provides a summary and comparison of regulations for adhesion quality in the countries mentioned above. While many factors contribute to EIFS adhesion quality, most regulations emphasize performance tests, such as insulation durability and bonding strength tests, among other criteria [23].

Therefore, this study proposes that the automatic adhesion quality-inspection model can measure the shape, such as ribbon width, number of dabs, distances of dabs, diameter of dabs, and the adhesive area, which are items in the EIFS specifications.

3. CV-Based Automatic Adhesion Quality Inspection Model for EIFS

3.1. Framework

The process of the proposed model is shown in Figure 2. The proposed model consists of three steps: the first is the segmentation step to extract the insulation and adhesive from the background. For this step, the deep-learning-based segmentation model is trained by using the collected and annotated dataset. The second is the image transformation step to align the tilted insulation image into a front-facing insulation image by using the perspective transformation. The third is the measurement step to measure the adhesive shape, such as ribbon width, distance between data and dab diameter, and the amount of adhesive on an insulation unit.

3.2. Segmentation Step

This section describes two main procedures in the instance segmentation step. First, it explains the preparation of training datasets for ribbons and dabs. Unlike other general objects, ribbons have the characteristics of a polygon with a hole, so the process of integrating them into a single ribbon is described. Second, it provides an explanation of the YOLOv8s-seg model, which was adopted for instance segmentation.

3.2.1. Annotation for Ribbon with a Hole and Dab

A deep-learning-based segmentation model needs to be trained on a large number of annotated datasets to improve inference performance. By using Roboflow [24], the dataset was labeled into two classes: ribbon and dab.

Ribbons are polygons with a hole in the middle, unlike typical instance segments such as dabs. However, most instance segmentation models do not support polygons with holes. Therefore, a process to modify a ribbon into a single polygon is necessary, as illustrated in Figure 3.

First, each ribbon was labeled twice: once for the inner ribbon and once for the outer ribbon. The outer ribbon was arranged in a clockwise direction, while the inner ribbon was arranged in a counterclockwise direction. A random point on the inner ribbon was selected, and the closest point on the outer ribbon was identified. Using these points, the outer and inner ribbon classes were merged into a single ribbon class.

3.2.2. Instance Segmentation Based on YOLOv8s-seg

The YOLOv8 series represents a notable evolution in deep-learning models for object detection. Developed by the creators of YOLOv5, YOLOv8 maintains a similar overall style but introduces significant improvements and optimizations. Notably, it employs a C2f structure in place of YOLOv5’s C3 structure in its backbone network, enhancing gradient flow while preserving its lightweight nature. Beyond object detection, YOLOv8 supports additional tasks such as object tracking, instance segmentation, image classification, and key point detection. Its head structure differs notably from YOLOv5 due to the implementation of a decoupled head structure, enhancing performance and task adaptability. For loss function calculation, YOLOv8 adopts the TaskAlignedAssigner positive sample assignment strategy [25] and introduces the distribution focal loss, which addresses class imbalance and focuses on challenging examples during training [26]. Training strategies include disabling mosaic augmentation in the final 10 epochs to improve precision in data augmentation, inspired by techniques from YOLOX [27].

YOLOv8-seg, an extension of YOLOv8, specializes in segmentation tasks, leveraging principles from YOLACT to achieve real-time instance segmentation with high segment mean average precision [28]. The YOLOv8-seg network is structured around two primary components: the backbone and the head, with the head further subdivided into neck and segment modules. This network series includes five distinct scale models: YOLOv8n-Seg, YOLOv8s-Seg, YOLOv8m-Seg, YOLOv8l-Seg, and YOLOv8x-Seg. Each model is designed to cater to specific computational capabilities and performance criteria, offering flexibility in various applications of object detection and segmentation tasks.

The YOLOv8s-seg model was chosen based on its segment mask [email protected] performance and lightweight design requirements for experimentation in this study. Figure 4 depicts the architecture of the YOLOv8s-seg network.

3.3. Image Transformation Step

Although the instance segmentation outputs are well performed, they still have limitations when inspecting adhesion quality. To accurately measure the shapes and areas of the adhesive, the images must be front faced to minimize distortion. This section describes the procedures for image transformation to align the tilted shapes into proper front-facing images.

3.3.1. Four-Vertices Point Extraction

To measure the shape and area of adhesive, the tilted insulation image needs to be aligned into a front-facing image. As a first step, four vertices that form the insulation on the image, along with their corresponding real-world coordinate pairs, are extracted. To identify these four points, the RANSAC (RANdom SAmple Consensus) algorithm [29], which is effective in accurately estimating model parameters even in the presence of significant outliers, was employed. Using this algorithm, four straight lines aligned with the edges of the ribbon mask were identified (Figure 5a). Next, based on the boundary and endpoints of each edge pair, the four vertices of the ribbon mask were determined, as illustrated in Figure 5b.

3.3.2. Perspective Transformation Step

The tilted insulation in an image is aligned into a front-facing insulation by using perspective transformation. A perspective transformation is the process of transforming an image so that it can be viewed from a different perspective or viewpoint. For perspective transformation, a transformation matrix, called a homography $H$, is generated using the pre-transformation image coordinates $(x, y)$ and post-transformation image coordinates $(x’, y’)$ and applied to the mask of ribbon and dab to convert the front-facing image, as shown in Equations (2) and (3).

(2)$\left[\begin{array}{c}{x}^{’}\\ y^{’}\\ z^{’}\end{array}\right]=H\left[\begin{array}{c}x\\ y\\ z\end{array}\right]$

(3) $H= \left[\begin{array}{ccc}{h}_{11}& h_{12}& h_{13}\\ h_{21}& h_{22}& h_{23}\\ h_{31}& h_{32}& h_{33}\end{array}\right]$

The warpPerspective method from opencv-python [30] is used for this purpose. The pre-transformation image coordinates are the four vertices extracted in Section 3.3.1, and the post-transformation image coordinates are set to [0, 1200], [600, 0], and [600, 1200], converting the insulation size of 600 mm $\times$ 1200 mm to 1 pixel per mm.

Through the image transformation step, the insulation in the image taken from different angles in the field can be fitted to the same resolution as the actual size of the insulation board (600 $\times$ 1200). The proposed model measures the shape and area of adhesive based on this image.

3.4. Measurement Step

The proposed model measures the ribbon width, dab diameter, dab distance, and adhesive (ribbon and dab) area. These inspection items are described in Figure 6, and the detailed measurement steps are explained in the following subsections.

3.4.1. Ribbon Width

The ribbon width is calculated by dividing it into horizontal ribbon (top and bottom edges) and vertical ribbon (left and right edges), as shown in Algorithm 1. First, a transformed ribbon mask is divided into four segments based on the middle of the insulation height and width. In each segment, the difference between minimum and maximum y-values at the same x-value is the horizontal ribbon width. Conversely, the difference between the minimum and maximum $x$-values at the same $y$-value in each segment represents the vertical ribbon width.

Algorithm 1. Measuring the ribbon width

1:   Input: transformed ribbon mask (600 × 1200)2:   split the transformed ribbon mask into four segments based on the middle of the height and width of the insulation size3:   for each segment in four segments4:     for$\mathrm{each} x$ value5:       $\mathrm{find} \mathrm{the} \mathrm{maximum} \mathrm{and} \mathrm{minimum} y$-values6:       $\mathrm{horizontal} \mathrm{ribbon} \mathrm{width}=(\mathrm{maximum} y$$-\mathrm{value}–\mathrm{minimum} y$-value)7:     for$\mathrm{each} y$ value8:       $\mathrm{find} \mathrm{the} \mathrm{maximum} \mathrm{and} \mathrm{minimum} x$-values 9:       $\mathrm{vertical} \mathrm{ribbon} \mathrm{width}=(\mathrm{maximum} x$$-\mathrm{value}–\mathrm{minimum} x$-value)10:   Output: horizontal and vertical ribbon widths

3.4.2. Dab Distance

The dab distance is measured as the distance between the centroids of two horizontally or vertically adjacent dabs, and the distance between the centroid and adjacent edges (Algorithm 2).

Firstly, to find a centroid of each dab, a clustering process for each dab is performed using the K-means clustering algorithm, which is an unsupervised learning algorithm for partitioning a dataset into $k$ clusters. The proposed model fixes the k value to 8. Secondly, the centroids of all dabs are sorted in ascending order for each of the $x$ and $y$ axes, and then they are assigned an index from top left to bottom right.

Algorithm 2. Preprocessing for dab distance calculation

1:   Input: all transformed dab masks2:   for each transformed dab mask 3:     $\mathrm{find} \mathrm{centroid} \mathrm{of} \mathrm{transformed} \mathrm{dab} \mathrm{mask} \mathrm{by} \mathrm{using} \mathrm{K}-\mathrm{mean} \mathrm{clustering} \mathrm{algorithm} (k$ = 8)4:     sort all centroids by ascending order of y-values5.   for$n\in \mathrm{1,2},\dots ,\mathrm{t}\mathrm{h}\mathrm{e}\mathrm{n}\mathrm{u}\mathrm{m}\mathrm{b}\mathrm{e}\mathrm{r}\mathrm{o}\mathrm{f}\mathrm{d}\mathrm{a}\mathrm{b}\mathrm{s}/ 2$ 6:     $\mathrm{sort} (2n$$)\mathrm{th} \mathrm{and} (2n+1$)th centroid points by ascending order of x-values7:   Output: indexed centroids of dabs

The distance between the two adjacent centroids of dabs $\left(c_{ix},c_{iy}\right) \left(i=1\dots 8\right)$ is calculated using the Euclidean distance shown in Equation (4).

(4)$centroid distance=\left\{\begin{array}{c}\sqrt{{\left(C_{ix}-C_{(i+1)x}\right)}^2+{\left(C_{iy}-C_{(i+1)y}\right)}^2}, i\in 1, 3, 5\\ \sqrt{{\left(C_{ix}-C_{(i+2)x}\right)}^2+{\left(C_{iy}-C_{(i+2)y}\right)}^2}, i\in 1, 2, 3, 4\end{array}\right.$

The distance between the centroid and adjacent edges computes the distance from the x, y coordinates of the centroids to the straight-line equation of the adjacent insulation edges ($x=0, x=600, y=0, y=1200$) (Equation (5))

(5)$edge distance=\left\{\begin{array}{c}{C}_{ix}, i\in 1, 3, 5, 7\\ 600-C_{ix}, i\in 2, 4, 6, 8\\ C_{iy}, i\in 1, 2\\ 1200-C_{iy}, i\in 7, 8\end{array}\right.$

3.4.3. Dab Diameter

The diameter of the dab is calculated from the width and height. The width of a dab is calculated as the difference between the maximum and minimum $x$-values in the dab. The height of a dab is calculated as the difference between the maximum and minimum $y$-values in the dab.

3.4.4. Adhesive Area

The area of ribbons and dabs is calculated by counting the number of pixels in each transformed mask. The transformed insulation image through the image transformation step has a resolution of 600 pixels $\times$ 1200 pixels, and the insulation size is 600 mm $\times$ 1200 mm, so they correspond to 1 mm per pixel. If the transformed image has pixels of width $a$ and height $b$, then the adhesive area is calculated by multiplying the sum of the pixel counts by the area coefficients below (Equation (6)).

(6)$area coefficient = {\displaystyle \frac{600}{a}}\times {\displaystyle \frac{1200}{b}}$

where $a$ is the number of horizontal pixels of a transformed insulation image and $b$ is the number of vertical pixels of a transformed insulation image.

4. Case Study

4.1. Overview

A case study was conducted to validate the proposed model. The validation of the proposed model is divided into two parts: the performance of the instance segmentation model and the performance of the proposed model. To validate the proposed model, an iPhone 11 mini, a commonly available smartphone, was utilized to capture 300 images of EIFS construction work at three sites in Nowon-gu, Seoul. This choice was made to verify the general applicability of the model using standard imaging devices, ensuring that it is not reliant on specialized or high-end equipment. By leveraging a widely accessible device, the study highlights the practicality and adaptability of the proposed approach in real-world scenarios. The workstation used for a case study was equipped with 32 GB of RAM, a Core i7-12700KF CPU, and an NVIDIA GeForce RTX 3080 Ti.

4.2. Performance of Instance Segmentation

4.2.1. Evaluation Metrics

To evaluate the segmentation performance, the mean average precision (mAP) metric was utilized. Average precision (AP) is the area under the precision-recall curve for a specific class.

(7)$Precision={\displaystyle \frac{TP}{TP+FP}}$

(8)$recall={\displaystyle \frac{TP}{TP+FN}}$

(9)$\mathrm{A}\mathrm{P}=\underset{n\to \mathrm{\infty }}{\mathit{lim}}{\sum }_{k=1}^n\left(R_k-R_{k-1}\right)P_k={\int }_0^{1}p\left(r\right)dr$

(10)$\mathrm{mAP}={\displaystyle \frac{1}{N}}\sum _{i=1}^{N}A{P}_i$

where $R_k$ represents the k-th point on the precision-recall curve’s horizontal axis, $P_k$ represents the vertical axis at k. True positive (TP) refers to the count of instances correctly predicted as true by the model, false positive (FP) denotes instances wrongly predicted as true while they are actually false, and false negative (FN) represents instances incorrectly predicted as false that are actually true.

Precision (Equation (7)) is the proportion of correct predictions among the results predicted by the model, while recall (Equation (8)) is the proportion of actual true instances that the model correctly predicts as true.

mAP (Equation (10)) is the average of these AP values across all classes. Specifically, we used mAP@[IoU = 50:95] (mAP50:95), which averages the precision scores at IoU (Equation (11)) thresholds ranging from 50% to 95% in 5% increments. This provides a comprehensive measure of the model’s accuracy across various levels of instance segmentation.

(11)$IoU={\displaystyle \frac{Area of Overlap}{Area of Union}}$

4.2.2. Training and Test

A total of 300 images were annotated in the way mentioned in Section 3.2.1. The dataset was split training: validation: test = 8:1:1. The hyperparameters of the EIFS segmentation model were set at 200 epochs, a learning rate of 0.01, and a batch size of 16. As a result, the best model showed mAP50:95 of 0.86 at the 188th epoch (Figure 7). Figure 8 shows a sample of the test results.

4.3. Performance of Proposed Model

In this section, the measurement results of our EIFS inspection model are presented. Various factors of ribbons and dabs were evaluated using error metrics and root mean square error (RMSE).

Ground truth values were established through manual annotations. To validate our model, EIFS samples were tested under various conditions of shape, width, and area. The average error and RMSE were calculated to explain the accuracy of each measurement, and error analysis was conducted for different conditions to identify specific patterns or trends.

All results are presented in tables for comparison. Additionally, figures, graphs, and scatter plots were provided for better visual comprehension.

4.3.1. Evaluation Metrics

The performance of the proposed model was analyzed for accuracy on ribbon width, dab diameter, dab distance, and adhesive (ribbon and dab) area by comparing the predicted value by the proposed model and the ground true. To validate the performance of the proposed model in various environments, the test was conducted with 30 images of 9 different types of adhesive shapes and ribbon thicknesses. For ground truth, there are limitations to directly measuring the irregular area and thickness of adhesive in the field. So, a photo was taken of each type from the front and annotated, and the four vertices points were extracted manually. The ground truth was obtained after performing the image transformation step (Section 3.3) and measurement step (Section 3.4) with these images and required information.

The accuracy of width, dab diameter, and adhesive (ribbon and dab) area are calculated by Equation (12):

(12)$\mathrm{Error}={\displaystyle \frac{1}{N}}\sum _{i=1}^N{\displaystyle \frac{\left|G{T}_i-Pre{d}_i\right|}{G{T}_i}}$

where $G{T}_i$ represents the ground truth of the $i$-th order, and $Pre{d}_i$ is the prediction of the $i$-th order of the sample. $N$ is the total number of samples.

The accuracy of the dab distance was validated by evaluating the accuracy of the dab location coordinates using RMSE, which is calculated using Equation (13). Unlike other measurements that assess a single variable at a time, dab distance requires measuring the $x$ and $y$ coordinates of the centroids of eight dabs in a single evaluation. Therefore, RMSE, which integrates these two variables into a single metric, was selected as the evaluation metric.

(13)$\mathrm{RMSE}=\sqrt{{\displaystyle \frac{1}{n}}\sum _{i=1}^n\left({\left(G{T}_{x_i}-Pre{d}_{x_i}\right)}^2+{\left(G{T}_{y_i}-Pre{d}_{y_i}\right)}^2\right)}$

where $G{T}_{xi}$ is the ground truth of the $x$-coordinate of the $i$-th order, $Pre{d}_{xi}$ is the prediction of the $x$-coordinate of the $i$-th order of the sample, $G{T}_{yi}$ is the ground truth of the $y$-coordinate of the $i$-th order, and $Pre{d}_{yi}$ is the prediction of the $y$-coordinate of the $i$-th order of the sample. $n$ is the total number of samples.

4.3.2. Ribbon Width

The total error for ribbon width is 9.42% (accuracy 90.58%), with the error for each type detailed in Table 3. Figure 9 compares the ribbon width values to the ground truth, where the x-axis represents the spread insulation edges, and the y-axis indicates the ribbon width, showing both ground truth and predicted values for each type. Type 8 has the highest error at 22.58%, as shown in Figure 6, where the graph highlights that significant errors occur in the thinner parts of the ribbon width.

4.3.3. Dab Distance

The total error of dab distance was evaluated based on the error in the dab centroid location, and RMSE was used as a metric. The RMSE of the dab location was 6.9 mm; the RMSE of each type is shown in Table 3. Figure 10 is a dab centroids location scatter graph. The errors of type 1 and type 5 are high, measuring at 12.95 mm and 14.64 mm, respectively, which is likely due to the difficulty in extracting the centroid location caused by the irregularity of the dab.

4.3.4. Dab Diameter

The total error for the dab diameter is 4.72% (accuracy 95.28%), with the error for each type detailed in Table 4. Figure 11 illustrates the ground truth and predicted values of dab diameter for each type, offering a clear comparison of the model’s performance. Each rectangle represents the bounding box of the dabs, providing a visual comparison of the predicted and actual dab diameters, which helps to identify discrepancies and assess overall accuracy.

4.3.5. Adhesive Area

The total error of the adhesive area is 6.42% (accuracy 93.58%), which consists of the ribbon (8.66%) and dab (3.48%), and the error of each type is shown in Table 5. Figure 12 shows the ground true and prediction values of adhesive area for each type. Type 8 has the highest error at 10.37%, which is likely due to the small size of ribbons and dabs, which tend to have larger error values even if the difference value is similar to the other types.

4.4. Result Analysis and Discussion

Table 6 summarizes the performance of the proposed model. The bold numbers indicate larger values than others. The types with the largest errors for each inspection item are Type 8 for the ribbon width and adhesive area, Type 5 for the dab distance (centroid location), and Type 4 for the dab diameter.

The analysis of the error of the proposed model revealed the following. First, the segmentation of thin ribbons was not performed properly. Type 8 and Type 9, which had the lowest accuracy in ribbon width and adhesive area measurements, were samples with ribbon thicknesses that differed significantly from other EIFS samples. Figure 13a shows the segmentation result of Type 8, while Figure 13b shows the result of Type 9. In Figure 13a, the ribbon at the bottom was detected as larger than the insulation size, causing errors during the image transformation process.

Second, the angle at which the insulation is photographed, combined with variations in adhesive thickness, significantly affects the accuracy of the model. Heavily tilted photos, such as those shown in Figure 14, result in distortions during the image transformation process, particularly for objects farther away from the camera. These distortions directly contribute to measurement errors.

As shown in Figure 14b and Figure 15, which depict the same EIFS image, it is evident that heavily tilted images experience distortion during the image transformation process. This distortion ultimately impacts the accuracy of the EIFS inspection results.

To ensure high performance when using the proposed model in the field, it is necessary to consider the following. First, when photographing insulation in the field, take pictures of the insulation from the front. This will reduce distortion from the camera lens and distortion due to adhesive thickness. Second, take pictures of the insulation from as close as possible. The larger the object in the image, the easier it is to extract the four vertices required for perspective transformation. This can lead to more accurate image alignment and, therefore, better measurement results. Finally, take photos with minimal shadows on the insulation. This will reduce the chances of incorrect segmentation results due to shadows.

The proposed model can be used to determine the quality of the EIFS adhesive by setting a threshold value for each metric. Figure 16 illustrates the quality assessment results based on the following metric thresholds: a vertical distance of 50 cm between two dabs, a horizontal distance between dabs, a ribbon width of 30 cm, a dab diameter of 10 cm, and an adhesive area of 40%. Items that did not meet these criteria were marked in red. The proposed model can derive inspection results in seconds with only the collected photos, which can help automate the existing labor-intensive and time-consuming inspection work and increase the productivity of inspection work.

5. Conclusions

This study proposed an automatic inspection model of EIFS using a single image taken from the construction site. The proposed model consists of the segmentation step to extract the insulation and adhesive from the background, the image transformation step to align the tilted insulation image into a front-facing insulation image by using the perspective transformation, and the measurement step to inspect the adhesive shape, including ribbon width, dab distance, dab diameter, and the amount of adhesive on the insulation.

The proposed model was validated through a case study. The validation of the proposed model is divided into two parts: the performance of the instance segmentation model and the performance of the proposed model. In the case of the instance segmentation model, the accuracy of the model trained with 240 images was mAP50:95 of 0.86. In the case of the proposed model, the error was 9.42% for ribbon width, 4.72% for dab diameter, 6.42% for adhesive area, and RMSE 6.9 mm for dab distance. As a result, the proposed model was found to be valid.

This study can contribute to the body of knowledge as follows. First, the proposed model can quickly and quantitatively inspect the shape and amount of the adhesive, which are inspection items in the EIFS specification. It can overcome the traditional inspection method’s limitations, which are time-consuming and labor-intensive work. Second, the effectiveness of applying CV to the adhesion quality of EIFS is proven. The CV-based model in this study can measure the adhesive shape, including ribbon width, dab distance, dab diameter, and the amount of adhesive on an insulation with a high accuracy. Finally, this study’s methodology can be used for measuring the size or area of other objects in the construction site from a single image. For example, it can be used to measure the finished area of a wall and the spacing between supports.

The limitations of this study have been discussed in the section on results analysis and discussion. First, heavily tilted images can lead to distortion during the image transformation step. Second, variations in ribbon thickness, such as thinner ribbons, can affect the accuracy of inspection results due to oversegmented outputs.

Additionally, the number of images used for training the instance segmentation model is relatively small and was collected from a limited number of sites. As a result, the proposed model cannot guarantee high performance in other environments.

For further studies, the performance of the proposed model will be improved by increasing its robustness under varied image conditions and by acquiring a larger dataset from diverse environments. This will enhance the accuracy and generalizability of the instance segmentation model.

Author Contributions

Conceptualization, T.K.; Methodology, M.K.; Software, S.Y.; Formal analysis, M.K.; Writing—original draft, M.K.; Visualization, S.Y.; Supervision, T.K.; Project administration, T.K.; Funding acquisition, T.K. All authors have read and agreed to the published version of the manuscript.

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Data Availability Statement

The raw data supporting the conclusions of this article will be made available by the authors on request.

Conflicts of Interest

The authors declare no conflict of interest.

Footnotes

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Figures and Tables

Figure 1. Specification of EIFS adhesive in Korea (a) LH 46030 (b) LHCS 41 42 00 15 (unit mm).

Figure 2. Framework of CV-based automatic adhesion quality-inspection model for EIFS.

Figure 3. Illustration of creating a single polygon with a hole by merging a clockwise outer ribbon and a counterclockwise inner ribbon. The merging line (red) connects the closest point on the outer ribbon to a randomly selected point on the inner ribbon, integrating the two boundaries into a cohesive shape. For better visibility, the red lines are exaggeratedly spaced in the figure.

Figure 4. Architecture of YOLOv8-seg.

Figure 5. Four-vertices point extraction: (a) identifying four straight lines aligned with the edges of the ribbon mask using RANSAC algorithm, (b) determining the four vertices of the ribbon mask based on the intersection points of the lines identified in (a).

Figure 6. Inspection items of proposed model.

Figure 7. Training of model.

Figure 8. Validation of YOLOv8s-seg.

Figure 9. Graph comparing ground truth and predicted ribbon width values.

Figure 10. Scatter plot for dab location.

Figure 11. Result of dab diameter.

Figure 12. Result of adhesive area.

Figure 13. Examples of thin ribbon (a) segmentation of Type 8 (b) segmentation of Type 9.

Figure 14. Examples of heavy tilt (a) segmentation of Type 1 (b) segmentation of Type 5.

Figure 15. Examples of thickness distortion: (a) the original image and (b) the aligned image after the image transformation process. It can be observed that heavily tilted images become distorted after the image transformation process, especially in regions farther from the camera.

Figure 16. EIFS adhesive quality determination based on threshold metrics.

Table 1

Specification of ETAG 004.

Mode of Failure	The Minimum Failure-Resistance Values After Each Conditioning in N/mm2 (MPa)
	Dry Condition	After Effect of Water
		At 2 h After Removing the Samples from the Water	At 7 Days After Removing the Samples from the Water
Adhesive rupture	0.08 *	Adhesive ruptureCohesive rupture in adhesive	0.08 *
Cohesive rupture in adhesive
Cohesive rupture in insulation	0.03	Cohesive rupture in insulation	0.03

* One single value lower than 0.08 but higher than 0.06 N/mm² (MPa) is admissible.

Table 2

Summary and comparison of regulations for adhesion quality.

Author	ASTM	EOTA	LH	LHCS
Area of adhesive	40%	0.03/Bwhere B is minimum failure resistance of the adhesive	40%	40%
Size of insulation board	48 in. (1219 mm) ± 1/16 in. (1.6 mm) × 24 in. ± 1/16 in. (1.6 mm).	X	1200 mm × 600 mm	1200 mm × 600 mm
Width of ribbon	2 in. (50.8 mm)	X	50 mm	50 mm
Number of dab	X	X	6	8
Distances of dab	No more than 8 in. (203.2 mm)	X	150 mm from board 450 mm between dabs	X
Diameter of dab	Not less than 4 in. (101.6 mm)	X	100 mm	100 mm
Figure	X	X	O	O

Table 3

Error of ribbon width and RMSE of dab distance.

Type	Ribbon Width	Dab Distance
	Error	RMSE
Type 1	7.55%	12.95 mm
Type 2	6.94%	4.13 mm
Type 3	5.70%	4.10 mm
Type 4	5.77%	6.16 mm
Type 5	9.29%	14.64 mm
Type 6	6.83%	3.95 mm
Type 7	7.97%	4.76 mm
Type 8	22.58%	7.17 mm
Type 9	12.11%	4.20 mm
Total	9.42%	6.90 mm

Table 4

Error of dab diameter.

Type	Dab 1	Dab 2	Dab 3	Dab 4	Dab 5	Dab 6	Dab 7	Dab 8	Total
Type 1	4.12%	6.58%	3.53%	6.10%	4.59%	3.31%	3.87%	6.70%	4.85%
Type 2	7.19%	4.51%	5.33%	4.17%	3.57%	5.26%	3.77%	4.68%	4.81%
Type 3	4.03%	3.87%	4.14%	5.95%	4.74%	2.17%	4.22%	2.36%	3.93%
Type 4	10.67%	7.08%	5.17%	7.31%	5.21%	7.20%	5.90%	13.19%	7.72%
Type 5	6.94%	4.38%	4.13%	3.47%	4.54%	3.89%	4.03%	4.49%	4.48%
Type 6	5.86%	3.85%	3.96%	8.98%	6.84%	6.36%	1.95%	4.03%	5.23%
Type 7	4.03%	4.18%	4.53%	5.12%	4.15%	3.57%	3.52%	2.88%	4.00%
Type 8	3.55%	3.54%	2.19%	2.32%	4.89%	2.47%	4.45%	5.51%	3.61%
Type 9	4.70%	5.09%	4.66%	4.63%	4.87%	4.38%	4.58%	2.16%	4.39%
Total									4.78%

Table 5

Error of adhesive area.

Type	Ribbon	Dab	Total
Type 1	9.00%	2.29%	6.12%
Type 2	7.24%	3.78%	5.81%
Type 3	8.85%	7.61%	8.34%
Type 4	8.47%	0.95%	5.39%
Type 5	1.56%	1.02%	1.23%
Type 6	6.21%	3.98%	5.35%
Type 7	9.91%	3.45%	7.14%
Type 8	15.33%	4.33%	10.37%
Type 9	11.37%	3.92%	8.01%
Total	8.66%	3.48%	6.42%

Table 6

Summary of the proposed model’s performance.

	Ribbon Width	Dab Distance	Dab Diameter	Adhesive Area
	(Ratio)	(RMSE)	(Ratio)	(Ratio)
Type 1	7.55%	12.95 mm	4.85%	6.12%
Type 2	6.94%	4.13 mm	4.81%	5.81%
Type 3	5.70%	4.10 mm	3.93%	8.34%
Type 4	5.77%	6.16 mm	7.72%	5.39%
Type 5	9.29%	14.64 mm	4.48%	1.23%
Type 6	6.83%	3.95 mm	5.23%	5.35%
Type 7	7.97%	4.76 mm	4.00%	7.14%
Type 8	22.58%	7.17 mm	3.61%	10.37%
Type 9	12.11%	4.20 mm	4.39%	8.01%
Total	9.42%	6.90 mm	4.78%	6.42%

The bolded values indicate relatively higher errors.

References

1. Ascione, F.; Bianco, N.; Mauro, G.M.; Napolitano, D.F. Building envelope design: Multi-objective optimization to minimize energy consumption, global cost and thermal discomfort. Application to different Italian climatic zones. Energy; 2019; 174, pp. 359-374. [DOI: https://dx.doi.org/10.1016/j.energy.2019.02.182]

2. Ferrara, M.; Virgone, J.; Fabrizio, E.; Kuznik, F.; Filippi, M. Modelling Zero Energy Buildings: Parametric Study for the Technical Optimization. Energy Procedia; 2014; 62, pp. 200-209. [DOI: https://dx.doi.org/10.1016/j.egypro.2014.12.381]

3. (The Goverment of the Republic of Korea). 2050 Carbon Neutral Strategy of the Republic of Korea Towards a Sustainable and Green Society; The Goverment of the Republic of Korea: Seoul, Republic of Korea, 2020.

4. Xu, H.; Wang, H.; Huo, Q.; Qin, Y.; Zhou, H. Comparative study of Chinese, European and ISO external thermal insulation composite system (ETICS) standards and technical recommendations. J. Build. Eng.; 2023; 68, [DOI: https://dx.doi.org/10.1016/j.jobe.2022.105687]

5. Evola, G.; Margani, G.; Marletta, L. Energy and cost evaluation of thermal bridge correction in Mediterranean climate. Energy Build.; 2011; 43, pp. 2385-2393. [DOI: https://dx.doi.org/10.1016/j.enbuild.2011.05.028]

6. Gonçalves, M.; Simões, N.; Serra, C.; Flores-Colen, I. A review of the challenges posed by the use of vacuum panels in external insulation finishing systems. Appl. Energy; 2020; 257, 114028. [DOI: https://dx.doi.org/10.1016/j.apenergy.2019.114028]

7. Seunghwan, W.; Sungwoong, Y.; Beom,; Sumin, K. Exterior insulation finishing system using cementitious plaster/microencapsulated phase change material for improving the building thermal storage performance. Constr. Build. Mater.; 2021; 299, 123932. [DOI: https://dx.doi.org/10.1016/j.conbuildmat.2021.123932]

8. Army Engineer District Omaha NE. Considerations for the Use of Exterior Insulation & Finish Systems (EIFS) on US Army Facilities; US Army Corps of Engineers, Construction Engineering Research Laboratory: Champaign, IL, USA, 1991.

9. Kuo-Liang, L.; Jhih-Long, F. Applications of computer vision on tile alignment inspection. Automat Constr.; 2013; 35, pp. 562-567. [DOI: https://dx.doi.org/10.1016/j.autcon.2013.01.009]

10. Silva, W.R.L.d.; Lucena, D.S.d. Concrete Cracks Detection Based on Deep Learning Image Classification. Proceedings; 2018; 2, 489. [DOI: https://dx.doi.org/10.3390/ICEM18-05387]

11. Dong, W.; Bingyu, R.; Bo, C.; Jiajun, W.; Xiaoling, W.; Tao, G. Real-time monitoring for vibration quality of fresh concrete using convolutional neural networks and IoT technology. Automat Constr.; 2021; 123, 103510. [DOI: https://dx.doi.org/10.1016/j.autcon.2020.103510]

12. Cha, Y.-J.; Choi, W.; Büyüköztürk, O. Deep Learning-Based Crack Damage Detection Using Convolutional Neural Networks. Comput. -Aided Civil. Infrastruct. Eng.; 2017; 32, pp. 361-378. [DOI: https://dx.doi.org/10.1111/mice.12263]

13. Dongho, J.; Min Kyoung, K.; Yeounung, J.; Jae Eun, O.; Juhyuk, M.; Dong Joo, K.; Seyoon, Y. High-accuracy rebar position detection using deep learning–based frequency-difference electrical resistance tomography. Automat Constr.; 2022; 135, 104116. [DOI: https://dx.doi.org/10.1016/j.autcon.2021.104116]

14. Nguyen Manh, T.; Yije, K.; Jung-Yoon, L.; Sangyoon, C. Automatic Stereo Vision-Based Inspection System for Particle Shape Analysis of Coarse Aggregates. J. Comput. Civil. Eng.; 2022; 36, 04021034. [DOI: https://dx.doi.org/10.1061/(ASCE)CP.1943-5487.0001005]

15. Ziyue, Z.; Yongqi, W.; Zhenhua, W.; Wu, Y.; Changying, W.; Bin, H.; Mingzi, G.; Jiansen, Y. Deep learning enabled particle analysis for quality assurance of construction materials. Automat Constr.; 2022; 140, 104374. [DOI: https://dx.doi.org/10.1016/j.autcon.2022.104374]

16. Vasquez, J.; Furuhata, T.; Shimada, K. Image-Enhanced U-Net: Optimizing Defect Detection in Window Frames for Construction Quality Inspection. Buildings; 2024; 14, 3. [DOI: https://dx.doi.org/10.3390/buildings14010003]

17. Yoon, S.; Kang, M.; Jang, H.; Kim, T. Computer Vision-based Automated Adhesive Quality Inspection Model of Exterior Insulation and Finishing System. J. Korea Inst. Build. Constr.; 2023; 23, pp. 165-173.

18. E2430/E2430M-19 A. Standard Specification for Expanded Polystyrene (“EPS”) Thermal Insulation Boards for Use in Exterior Insulation and Finish Systems (“EIFS”); Technical Report ASTM International: West Conshohocken, PA, USA, 2018.

19. C1397-13 A. Standard Practice for Application of Class PB Exterior Insulation and Finish Systems (EIFS) and EIFS with Drainage; Technical Report ASTM International: West Conshohocken, PA, USA, 2019.

20. ETAG004 Guideline for European Technical Approval of External Thermal Insulation Composite Systems (Etics) with Rendering; European Organisation for Technical Approvals: Brussels, Belgium, 2013.

21. LH46030 LH Guide Specifications for Housing Projects Section 46030; Korea Land & Housing Corporation: Jinju-si, Republic of Korea, 2012.

22. LHCS 41 42 00 15 Korea Land& Housing Corporation Construction Specification LHCS 41 42 00 15; Korea Land & Housing Corporation: Jinju-si, Republic of Korea, 2020.

23. Franzoni, E.; Pigino, B.; Graziani, G.; Lucchese, C.; Fregni, A. A new prefabricated external thermal insulation composite board with ceramic finishing for buildings retrofitting. Mater. Struct.; 2015; 49, pp. 1527-1542. [DOI: https://dx.doi.org/10.1617/s11527-015-0593-7]

24. Dwyer, B.; Nelson, J.; Hansen, T. Roboflow (Version 1.0) [Software]. Available online: https://roboflow.com (accessed on 2 December 2024).

25. Feng, C.; Zhong, Y.; Gao, Y.; Scott, M.R.; Huang, W. TOOD: Task-aligned One-stage Object Detection. Proceedings of the 2021 IEEE/CVF International Conference on Computer Vision (ICCV); Montreal, BC, Canada, 11–17 October 2021; pp. 3490-3499. [DOI: https://dx.doi.org/10.1109/ICCV48922.2021.00349]

26. Li, X.; Wang, W.; Wu, L.; Chen, S.; Hu, X.; Li, J.; Tang, J.; Yang, J. Generalized Focal Loss: Learning Qualified and Distributed Bounding Boxes for Dense Object Detection. NIPS’20: Proceedings of the 34th International Conference on Neural Information Processing Systems, Vancouver, BC, Canada, 6–12 December 2020; Larochelle, H. Curran Associates Inc.: Red Hook, NY, USA, 2020; pp. 21002-21012.

27. Ge, Z.; Liu, S.; Wang, F.; Li, Z.; Sun, J. Yolox: Exceeding yolo series in 2021. arXiv; 2021; arXiv: 2107.08430

28. Yue, X.; Qi, K.; Na, X.; Zhang, Y.; Liu, Y.; Liu, C. Improved YOLOv8-Seg Network for Instance Segmentation of Healthy and Diseased Tomato Plants in the Growth Stage. Agriculture; 2023; 13, 1643. [DOI: https://dx.doi.org/10.3390/agriculture13081643]

29. Fischler, M.A.; Bolles, R.C. Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM; 1981; 24, pp. 381-395. [DOI: https://dx.doi.org/10.1145/358669.358692]

30. Bradski, G. The OpenCV Library. Dr. Dobb’s J. Softw. Tools; 2000; 25, pp. 120-125.