Architects believe that FFC reveals the most essential architectural aesthetics (Coburn et al., 2019). Compared with ordinary concrete, FFC has been widely used in domestic and international buildings for its natural surface as a decorative surface, which has a high-quality appearance while satisfying mechanics and durability. The decorative surface of FFC reduces the increased costs associated with additional decorative surfaces, so the probability of budget changes during construction is low (Miranda et al., 2020). FFC has been applied and achieved positive results not only in transportation centers and commercial buildings such as Chongqing West Station, Raffles Chengdu and Raffles Chongqing, but also in modern religious buildings (Ribeiro and Xavier, 2019). The chromatic aberration control on the FFC decorative surface affects the overall effect of the building, so chromatic aberration is a key point in the control of FFC appearance. At present, the research on the appearance quality of FFC is divided into two main directions, one is to study the factors affecting chromatic aberration and their improvements, and the other is to study the evaluation method of the FFC appearance quality.

The appearance quality of FFC is directly influenced by its own factors such as mix ratio and saturation. Studies have shown that the increase in cement saturation is beneficial to the appearance quality of FFC, but the mechanical properties of concrete decrease (Ajtayne Karolyfi and Papp, 2021). The self-compaction and appearance quality of FFC can be improved via adjusting the coarse aggregate gradation, but its application is limited due to the complex processes and high costs (Wu et al., 2021). The surface color and brightness of FFC are influenced by the cement content, water-cement ratio and sand ratio in that order (Mohammed et al., 2021). In the field of FFC apparent quality identification, there are more studies on crack identification and bubble identification. The diagnosis and repair of FFC surfaces can be performed by collecting images of the original FFC structure and combining them with machine vision techniques (Piferi, 2018), but there are still irreparable defects in the concrete appearance quality, and the diagnosis and repair work will be resource-intensive. This manuscript designs an experimental data acquisition method for FFC surface chromatic aberration based on UAV technology, it can rapidly acquire large area sample images, reduce the interference factors of image data, and improve the efficiency and quality of data acquisition. In addition, this method reduces the risk factor for staff.

The appearance quality of FFC is also indirectly influenced by factors in the production environment. Studies have shown that the effects of acid rain, haze, freeze-thaw and carbonation can cause significant chromatic aberrations in the facade of FFC buildings (Zhang et al., 2013). The dark chromatic aberration will destroy the surface appearance of FFC, and the dark chromatic aberration is more serious in the production process of outdoor FFC in winter (Chang et al., 2021). The process of removing the surface coating of FFC and mismatched release agents can lead to defects in the FFC appearance quality (Strehlein and Schiessl, 2008; Shyha et al., 2016). FFC buildings produce new chromatic aberrations in use, so it is necessary to study the identification and quantification of FFC chromatic aberration. The existing studies on the FFC surface CAI are few and cannot effectively quantify the FFC chromatic aberration.

Image processing technique can provide objective results quickly (Abouhamad et al., 2017; Mohan and Poobal, 2018; Payab et al., 2019; Chen et al., 2019; Tang et al., 2019). The identification and quantification of FFC by multi-intelligent body fuzzy systems (Payab and Khanzadi, 2021), the defects identification in the FFC appearance quality by laser point thermography (Scalbi et al., 2019), and the recording of the cracks beginning and development in combination with timelines (Prasanna et al., 2016) have been implemented. Scholars have identified the bubble contours on the FFC surface by extracting the grayscale values of the concrete surface images (Liu and Yang, 2017; Yoshitake et al., 2018). Scholars avoid the influence of subjectivity by predicting concrete performance through machine learning methods (Feng et al., 2022). This manuscript builds a FFC chromatic aberration identification model based on the optimized MSLSDA, avoids the problem of subjectivity in image processing techniques.

At present, the evaluation of the appearance quality of FFC is mainly through the image processing method (Liu et al., 2019), and the image data is segmented by selecting the appropriate threshold value (Yao et al., 2021). Therefore, the image processing process can only be applied to the current experimental data and lacks the ability to generalize to similar data (Li et al., 2020). Unlike cracks and bubbles on the surface of FFC, chromatic aberrations do not have clear boundaries and a fixed range of grayscale values, which cannot be quantified directly. This manuscript implements the FFC chromatic aberration quantization function through color space conversion, histogram equalization, image color recognition, image noise reduction and image mask operation.

A MSLSDA for FFC surface CAI is proposed in this manuscript. Firstly, the problem of subjectivity in CAI is solved by learning chromatic aberration features and classifying chromatic aberration by MSLSDA. Then, the surface chromatic aberration of FFC is quantitatively segmented by means of image processing, boundary features, geometric parameter features (length and width), chromatic aberration ratio features, total chromatic aberration ratio and number of chromatic aberration are calculated. Finally, the FFC-CADP is built to realize the prediction function and quantitative segmentation function of the FFC chromatic aberration.

FFC buildings often have a large project volume, making it difficult for inspectors to take samples and test them. Therefore, UAV is used for image acquisition in this study (Siebert and Teizer, 2014), as shown in Figure 1. The China Overseas HuanYu Art Center and Rail Transit R1 in Jinan are identified as sample buildings for experimental data through fieldwork, as shown in Figure 2. The image acquisition device is DJ Mavic two Pro, and the performance parameters are shown in Table 1. UAV has high vertical hovering accuracy and horizontal hovering accuracy, L1D-20c Hasselblad camera lens angle can reach 84°, the nearest focus point 1m, the lens can achieve autofocus, can meet the observation requirements of 5 m from the wall. The photo resolution can reach 5,472*3,684, and the high pixel can meet the requirements of the experiment on the data amount. Positive endurance and stable flight speed can ensure the efficiency of image acquisition.

According to the current Technical Specification for Fair-faced Concrete Construction JGJ169-2009, the identification method for the FFC appearance quality shall conform to the provisions of Table 2 (Kang and Cha, 2018). To ensure the consistency of the experimental data, the UAV is 5.0 m away from the target plane, the lens normal direction is always perpendicular to the target plane, and the shooting environment is naturally illuminated under normal weather. The combination of shooting angle and plane to be photographed include vertical flat surfac, vertical curved surface, non-vertical flat surfac and non-vertical curved surface. Currently, this study addresses the chromatic aberration identification in the FFC vertical flat surfac. The cases of vertical curved surface, non-vertical flat surfac and non-vertical curved surface will be considered in further studies to improve the generalizability of the FFC chromatic aberration identification method.

The original image is taken by UAV at a distance of 5 m from the FFC vertical flat surfac. Obstructions obscuring light rarely occurs during the original images acquisition. And we consciously avoid shooting original images with obstructions obscuring them. In the process of dataset production, we carefully screen and remove the sample images with obstructions obscuring them. However, the phenomenon of obstructions obscuring light and creating shadow area on the FFC does exist. The pixel value of the shadow area is affected by the light intensity and obstruction position, which we will try to solve in further studies.

In order to test the ability of the MSLSDA to generalize the FFC chromatic aberration problem, chromatic aberration data were collected from China Overseas HuanYu Art Center and Rail Transit R1 during the period of 12:30–14:30 on 22 March 2021, with 13:10–14:00 as the site transition period. China Overseas HuanYu Art Center is the main sample building, and Rail Transit R1 is the sub-sample building. According to the sample buildings, 410 original images of the main sample buildings are used for the training set and validation set, and 102 original images of the sub-sample buildings are used for the generalized prediction set (Cha et al., 2017).

In order to improve the training efficiency of MSLSDA, the original data is pre-processed in this study. The original image with a resolution of 5,472 × 3,684 is segmented into 306 sample images with a resolution of 256 × 256. After excluding the images with non-chromatic appearance quality indicators from the segmented sample images, a total of 2,368 experimental sample images are obtained, as shown in Figure 3.

Among them, 1894 experimental sample images are selected from 410 original images of the main sample buildings, accounting for 80% of the total number of experimental sample images. The 474 experimental sample images are selected from the original images of 102 sub-sample buildings, accounting for 20% of the total number of experimental sample images. The 474 experimental sample images include 237 sample images without chromatic aberration and 237 sample images with chromatic aberration. 80% of the 1894 experimental sample images are randomly selected as the training set and the remaining 20% as the validation set. The validation set consists of 189 sample images without chromatic aberration and 189 sample images with chromatic aberration. The training set consists of 758 sample images without chromatic aberration and 758 sample images with chromatic aberration.

In this manuscript, generalization prediction experiments are designed to enhance the generalization ability of the MSLSDA for the FFC surface CAI problem. The optimized and trained MSLSDA will sequentially predict each experimental sample image in the generalized prediction set and output the identification result and probability value prod. The identification results are compared with the actual results to quantitatively evaluate the generalization prediction ability of the MSLSDA. The probability value prod can quantitatively evaluate the probabilistic reliability of the identification results.

Labeling and path standardization of the experimental samples are crucial to the process of producing the dataset. Firstly, input the experimental sample images and save the structured data into fixed channels and fixed dimensions. Then, the dataset form is saved as Numpy form and the dataset is reordered. Finally, the 2,368 experimental sample images in the dataset are written to the corresponding paths in the images folder to complete the dataset.

Residual Network CNN’s lightweight residual network structure allows for deeper learning of more features layer by layer, making the forward and backward propagation process smooth, and it is more suitable for mobile terminal environments. Therefore, the Residual Network CNN lightweight structured data algorithm framework is used to identify the surface chromatic aberration of FFC. The Residual Network CNN framework accomplishes two main tasks, including qualitative and confidence evaluation of classification accuracy on structured data. The goal of structured data qualitative is to classify the color status of sample images. Confidence is a criterion to judge whether the classification results satisfy the probability requirement or not.

In order to test the generalization ability of the MSLSDA for FFC surface CAI, 474 sample images of sub-sample buildings are performed full-sample identification. The output sample images include light color without chromatic aberration (LW) output, dark color without chromatic aberration (DW) output, light color complete chromatic aberration (LCC) output, light color incomplete chromatic aberration (LIC) output, dark color complete chromatic aberration (DCC) output and dark color incomplete chromatic aberration (DIC) output, as shown in Figure 11. The first index is the result of CAI, abnomal indicates chromatic aberration in sample image and nomal indicates uniform color of the sample image. The second index indicates the probability value prod of the target dimension, which illustrates the confidence of the CAI results.

The generalized prediction set contains four defect categories of light FFC, dark FFC, complete chromatic aberration and incomplete chromatic aberration. And it contains four comprehensive defect categories of LCC sample images, DCC sample images, LIC sample images and DIC sample images, but the MSLSDA is still able to accomplish the CAI task. The two indexes show positive performance of FFC surface CAI. To represent the results of the generalization prediction experiments, the confidence data for the chromatic aberration and without chromatic aberration samples are labeled in Figure 12. It can be seen that the confidence range of the generalized prediction samples with chromatic aberration is almost between 0.88 and 1.00, and the confidence range of the generalized prediction samples without chromatic aberration is almost between 0.86 and 1.00. The accuracy of the chromatic aberration generalization prediction reaches 99.6%, which reflects the positive generalization prediction ability of the MSLSDA.

Through color space conversion, histogram equalization, image color recognition, image noise reduction and image mask operation, the multi-scale feature fusion of structured data is realized. The boundary features, geometric parameter features (length and width) and chromatic aberration ratio features of structured data are identified, as shown in Figure 13.

In order to perform histogram equalization of color sample images, the RGB color space of the structured data (shown in Figure 13A) is converted into an HSV color space that enables efficient image processing. Then the Void InRange function is used to identification whether each pixel point of src image is between lowerb value and upperb value. The pixel points between lowerb value and upperb value are set to 255, and the rest of the pixel points are set to 0. The pixel values are finally saved in the dst image, which gives a binary image corresponding to the target color, as shown in Figure 13B.

To remove the pretzel noise points in the binary image, a median filter method is used, as shown in Figure 13C. To remove the scattered distribution of noisy pixel points from the binary image, an open operation of erosion followed by expansion is used, as shown in Figure 13D. In order to reduce the independence of chromatic aberration areas and enhance the connectivity of individual chromatic aberration blocks, a closed operation of expansion followed by erosion is used, as shown in Figure 13E. In order to extract structural features of chromatic aberration in sample images, similarity variables are used to detect structural features similar to image masks, as shown in Figure 13F.

As can be seen in Figure 14, the sample image can be calculated by the masking algorithm and output the chromatic aberration mask index. The dimensional pixel values of the FFC chromatic aberration are more informative than the millimeter values during the inspection. Therefore, the mask algorithm trained in this manuscript directly outputs the pixel values of the chromatic aberration size. The index contains the chromatic aberration block number, chromatic aberration ratio features and geometric parameter features (length and width), these three parts listed in order from left to right. The chromatic aberration block number is based on chromatic aberration ratio features, and each chromatic aberration block is coded in order from the largest to the smallest, which is helpful for the inspector to quickly determine the largest chromatic aberration block. The chromatic aberration ratio features represent the proportion of a single chromatic aberration to the sample image, and further the total chromatic aberration ratio of the whole sample image can be obtained, which provides an accurate quantitative evaluation basis for inspectors. The geometric parameter features (length and width) are also expressed as pixel values in parentheses at the end of the index, the chromatic aberration block length pixel value on the right and the width pixel value on the left, which provides the inspector with information on the degree of chromatic aberration coverage.

As shown in Figure 15, all chromatic aberration blocks in the figure are identified and quantified. It can be seen that the Residual Network CNN framework and the chromatic aberration mask algorithm have positive identification and quantization performance not only for LCC sample images (shown in Figure 15A), but also for LIC sample images (shown in Figure 15B), DCC sample images (shown in Figure 15C) and DIC sample images (shown in Figure 15 (d)).

In order to illustrate the results of FFC identification and quantification in different states, The information of identification and quantification for different classes of experimental sample images in Figure 15 is listed in Table 4.

The FFC-CADP (shown in Figure 16) implements an automated procedure for identification and quantification of chromatic blocks on the surface image of FFC. The hardware environment for developing platform includes Core I7 8750h, 6 GB of gtx1060 and 16 GB of RAM. The operating system used to develop platform is Windows 10 × 64. The environment used to develop platform is Anaconda3 and Pycharm. To improve the applicability of platform, it is coded to apply to Windows 10 × 64 and ubuntu 18.04 × 64 operating system. The supporting environment for platform operation includs Pytorch 1.7.1, Opencv4, Numpy and Matplotlib. This study builds platform through Python 3.7 programming language.

Click the input button to switch the interface to the images folder path, and select the sample images to be identified and quantified to complete the input operation. The sample image after input is displayed in the first output box in the first line for comparison with the quantized output results. Click the projections button, the CAI result and prod will be output below the input sample image. Click the quantitative segmentation button to get color space recognition output, active noise reduction output, corrosion-expansion output, data closure operations output and quantitative segmentation output in sequence. The quantitative segmentation output contains the boundary features, geometric parameter features (length and width), chromatic aberration ratio features, total chromatic aberration ratio and number of chromatic aberration.

According to the actual working environment of FFC inspectors and combined with the demand of FFC surface CAI, the experimental data acquisition method of FFC surface chromatic aberration based on UAV technology is designed. Compared with traditional data acquisition methods, this method can quickly collect sample images in a large area, and reduce the interference factors in image data. Besides it can reduce the risk factor of staff, improve the efficiency and quality of data acquisition.

Based on the optimized MSLSDA, a FFC chromatic aberration identification model is built. The model can complete the quantization of FFC chromatic aberration quickly and accurately, and the quantization accuracy of the target reaches 92.1%.

The FFC-CADP is built to visualize the operation of identification and quantization on the FFC surface sample images. The calculations of the platform include image category return value, prod, boundary features, geometric parameter features (length and width), chromatic aberration ratio features, total chromatic aberration ratio and number of chromatic aberration.

Aiming at the problems of low efficiency and strong subjectivity in the quality inspection of FFC surface CAI, a method based on MSLSDA for FFC surface CAI is proposed. The method achieves the FFC chromatic aberration prediction function by FFC chromatic aberration identification model. The FFC chromatic aberration quantization function achieves by color space conversion, histogram equalization, image color recognition, image noise reduction and image mask operation.

This project will further investigate the performance and generalization prediction ability of different weight structured data algorithms for the FFC chromatic aberration dataset, and obtain the FFC surface CAI models suitable for different application scenarios.