KR20220156432A - Photo-based building construction defect analysis apparatus using deep learning - Google Patents

Abstract

The present invention relates to a photo-based building construction defect detection system using deep learning, and more specifically, to a photo-based building construction defect detection system using deep learning, which is to perform defect detection on an interior space after completion of building construction and to increase the defect detection performance. The photo-based building construction defect detection system comprises an interface unit and a defect detection unit.

Description

Photo-based construction defect detection system using deep learning {PHOTO-BASED BUILDING CONSTRUCTION DEFECT ANALYSIS APPARATUS USING DEEP LEARNING}

The present invention relates to a photo-based construction construction defect detection system using deep learning, and more particularly, to a photo-based construction using deep learning that detects defects in an indoor space where construction work is completed and improves defect detection performance. It is about a construction defect detection system.

Recently, as the fourth industrial revolution has developed significantly, computer-based automation technology has been in the limelight.

For example, artificial intelligence technology that learns what a person does by inputting specific data or rules into a machine or computer, and learning to recognize or decide a specific target by learning thousands or tens of thousands of big data artificial intelligence There is a deep learning technology that trains a computer to enable more complex thinking and judgment using a machine learning technology and an artificial neural network.

Recently, with the great development of computing speed and deep learning technology, it has been developed to the extent that it is possible to recognize or judge a specific object in a complex environment as well as simply recognizing a specific object.

Meanwhile, in order to recognize a specific object in an image, a boundary of the image is analyzed from the learned data, and a process of analyzing whether or not the analyzed boundary is a specific object is performed. Here, the more complex elements are included in the image, the more complex algorithm is required in the process of analyzing the boundary of the image and identifying the object.

In other words, more complex algorithms and more training data are needed to increase the accuracy of image identification.

However, in the process of learning such data, if a new object that has not been previously learned is included, there is a problem of not properly recognizing it.

In order to solve this problem, Patent Document 1 uses a damage detection system using AI to photograph an object by zooming a camera to a specific area at a facility inspection site and to find defects in the photographed image, but this is also expanded. When the processed image includes unlearned data, there is a problem in that the discriminative power of defects is reduced.

In addition, Patent Document 2 describes an augmented reality-based construction management system and a construction management method using the augmented reality implementation technology in an indoor or outdoor space, a digitized 3D building model storage and matching technology, and a construction manager's decision support technology. , including automatic recording of the building construction process, etc., since the scheduled air BIM model corresponding to the actual shooting screen is displayed on the screen of the mobile terminal, it is possible to check the completed and unfinished parts.

In addition, Patent Document 3 discloses a method for determining space state information of a building based on machine learning and a space state information determination server using the same, which complies with the current state of construction and safety rules based on subjective criteria of construction site officials or building officials subject to maintenance. Objectively discriminate spatial state information based on data so as not to erroneously determine whether or not there is a defect or whether a defect has occurred.

Registered Patent Publication No. 10-2260056 Registered Patent Publication No. 10-2162028 Registered Patent Publication No. 10-2298260

The present invention was invented to solve the problems of the prior art as described above, and more specifically, photo-based construction using deep learning that detects defects in interior images of construction works by applying deep learning defect classification technology instead of object detection. It is an object of the present invention to provide a construction defect detection system.

In addition, the present invention is to provide a photo-based construction construction defect detection system using deep learning that applies the condition of the learning image in consideration of the effect of each defect type by checking the minimum resolution of the image, the degree of influence of surrounding obstacles, and the degree of shading. for another purpose.

In addition, another object of the present invention is to provide a photo-based construction defect detection system using deep learning that processes construction verification by performing defect detection on images collected by an autonomous robot autonomously traveling in an indoor space. to be

The photo-based construction defect detection system using preferred deep learning of the present invention processes user manipulation and display user interface, transmits image selection and menu selection user manipulation to the defect recognition unit 30, and an interface unit 20 that displays image segmentation, defect detection, and detection results, which are operational states of the defect recognition unit 30, on a display; and a defect recognition unit including a still cut unit 31, an image segmentation unit 32, an analysis unit 33, and an output unit 34 that detects defects in an image of an indoor space where the construction work is completed ( 30); characterized in that it includes.

In addition, the defect recognizing unit 30 makes each still cut from the image selected by the interface unit 20, and divides the still cut into the image division unit ( 32) to the steel cut portion 31; an image segmentation unit 32 that divides the still cut at a constant ratio horizontally and vertically, and sequentially transmits the divided images to the analysis unit 33 so that defects are detected on the divided images; an analysis unit 33 which performs defect detection using a neural network on the image segmented by the image segmentation unit 32 and transmits a detection result to the output unit 34; And an output unit 34 that transmits the defect probability value, which is the detection result of the analysis unit 33, to the interface unit 20 so that the numerical output of the interface unit 20 and the defect confirmation procedure are performed. characterized by

In addition, the self-driving robot 40 further includes an autonomous driving unit 50 for autonomously traveling in an indoor space for construction work, photographing the indoor space, and transmitting the captured image to the defect recognition unit 30, The defect recognition unit 30 performs defect detection on the received image and transmits the detection result to the terminal 6, and the terminal 6 displays the defect detection result of the self-driving robot 40 on the screen. It is characterized by displaying in.

In addition, the analysis unit 33, the step of inputting a defect image (S101); Establishing an initial defect detection model for image segmentation and defect detection (S102); Determining defects in the divided images (S103); Secondarily classifying the image having a defect detection accuracy of 30 to 80% in step S103 (S104); Manually determining defects in the secondary classified images (S105); Collecting data of the image having a defect detection accuracy of 90% or more in the step S103 and the image determined to be defective in the step S105 (S106); monitoring the amount of learning of the defect detection model (S107); and securing deep learning learning data and building a defect detection model (S108).

In addition, the analysis unit 33, the step of starting the paper tear defect reading test (S201); Classifying the type of learning data (S202); Collecting defect-only images and foreground images (S203); Classifying the resolution (S204); Collecting 128, 245, 1024, 4096 size images (S205); Classifying the obstacle ratio (S206); Collecting images with obstacle ratios of 10%, 20%, 30%, and 40% (S207); Classifying defect types (S208); and collecting defect types A, B, C, and D (S209).

On the other hand, since the description of the technology disclosed in this specification is merely an embodiment for structural or functional description, the scope of rights of the disclosed technology should not be construed as being limited by the embodiment described in the text. That is, since the embodiment can be changed in various ways and can have various forms, it should be understood that the scope of rights of the disclosed technology includes equivalents capable of realizing the technical idea. In addition, since the object or effect presented in the disclosed technology does not mean that a specific embodiment should include all of them or only such effects, the scope of rights of the disclosed technology should not be construed as being limited thereto.

In addition, the meaning of terms described in the present invention should be understood as follows. Terms such as “first” and “second” are used to distinguish one component from another, and the scope of rights should not be limited by these terms. For example, a first element may be termed a second element, and similarly, a second element may be termed a first element.

Furthermore, it should be understood that when a component is referred to as being “connected” to another component, it may be directly connected to the other component, but other components may exist in the middle. On the other hand, when an element is referred to as being “directly connected” to another element, it should be understood that no intervening element exists. Meanwhile, other expressions describing the relationship between components, such as “between” and “between” or “adjacent to” and “directly adjacent to” should be interpreted similarly.

Singular expressions should be understood to include plural expressions unless the context clearly dictates otherwise, and terms such as “comprise” or “have” refer to the described feature, number, step, operation, component, part, or It should be understood that it is intended to indicate that a combination exists, and does not preclude the possibility of the presence or addition of one or more other features, numbers, steps, operations, components, parts, or combinations thereof.

The present invention can have the effect of performing interior finish verification after the construction work is completed by detecting defects in the interior image of the construction work by applying the deep learning defect classification technology instead of object detection.

In addition, the present invention can have an effect of improving defect detection performance by checking the minimum resolution of the image, the influence of surrounding obstacles, and the degree of shading and applying the conditions of the learning image in consideration of the effect of each defect type.

In addition, the present invention can have an effect of automating the defect detection task by performing defect detection on images collected by the self-driving robot autonomously traveling in an indoor space and processing construction work verification.

1 is an exemplary view illustrating a system authentication configuration that grants authority to execute hardware resources, an operating system, a core control unit, and a control unit operation for explaining the configuration of a photo-based construction defect detection system using deep learning of the present invention. to be.
FIG. 2 is a flow chart showing defect data collection and defect detection model construction used in the picture-based construction defect detection system using deep learning in FIG. 1 .
FIG. 3 is a flowchart illustrating a defect data reading test of the photo-based construction defect detection system using deep learning in FIG. 1 .
FIG. 4 is an exemplary view showing a user interface of the picture-based construction defect detection system using deep learning in FIG. 1 .
5 is an exemplary view of defect data that can be used in the picture-based construction defect detection system using deep learning in FIG. 1 .
6 is an exemplary diagram showing image segmentation, defect detection, and detection results of the photo-based construction defect detection system using deep learning in FIG. 1 .
FIG. 7 is an exemplary view showing detection results in which image segmentation and defect detection are applied to multiple indoor photos of the photo-based construction work defect detection system using deep learning in FIG. 1 .
8 is a block diagram showing the configuration of an autonomous robot to which the picture-based construction defect detection system using deep learning in FIG. 1 is applied.

Hereinafter, a photo-based construction defect detection system using deep learning according to a preferred embodiment of the present invention will be described in detail with reference to the drawings. Hereinafter, descriptions of conventionally known matters are omitted or simplified to clarify the gist of the present invention. The configurations included in the description of the present invention operate in individual or complex combination configurations.

1 is an exemplary view illustrating a system authentication configuration that grants authority to execute hardware resources, an operating system, a core control unit, and a control unit operation for explaining the configuration of a photo-based construction defect detection system using deep learning of the present invention. As such, referring to FIG. 1, the present invention includes a processor 1, a memory 2, an input/output device 3, a camera 10, an operating system 4, a control unit 5, an interface unit 20, and defect recognition. It includes part 30.

The processor 1 is a CPU (Central Processing Units), a GPU (Graphic Processing Unit), an FPGA (Field Programmable Gate Array), and an NPU (Neural Processing Unit). ) executes the executable code.

The memory 2 includes a permanent mass storage device such as random access memory (RAM), read only memory (ROM), disk drive, solid state drive (SSD), flash memory, etc. can do.

The input/output device 3 is an input device, such as a camera, keyboard, microphone, mouse, etc., including an audio sensor and/or an image sensor, and an output device, such as a display, speaker, haptic feedback device, etc. device may be included.

The camera 10 uses a Charge-Coupled Device (CCD) or a Complementary Metal-Oxide Semiconductor (CMOS) as an imaging device to convert light into an electronic signal and store it in the memory 2 . A large-capacity RAW image is stored in the memory 2, and the operating system 4 provides the RAW image in the memory 2 to the defect recognition unit 30.

The operating system 4 may include Windows, Linux, IOS, a virtual machine, a web browser, and an interpreter.

The control unit 5 determines the state of the input/output device 3 by sensor, key, touch, and mouse input under the support of the operating system 4, and performs an operation according to the determined state. The control unit 5 performs task scheduling using a timer and thread as a parallel execution routine.

The control unit 5 determines a state using the sensor value of the input/output device 3 and performs an algorithm according to the determined state.

The interface unit 20 processes a user interface such as user manipulation and display display, transmits user manipulations such as image selection and menu selection to the defect recognition unit 30, and divides images, which are operational states of the defect recognition unit 30, Defect detection and detection results are displayed on the display.

The defect recognition unit 30 performs defect detection on the image of the indoor space where the building work is completed, and uses the still cut unit 31, the image segmentation unit 32, the analysis unit 33, and the output unit 34. include

The still cut unit 31 creates each still cut from the image selected by the interface unit 20 and transfers the still cut to the image segmentation unit 32 so that image segmentation and defect detection are performed on one still cut.

The image segmentation unit 32 divides the still cut into a predetermined ratio horizontally and vertically, and sequentially transmits the divided images to the analysis unit 33 so that defects are detected on the divided images.

The analysis unit 33 performs defect detection using a neural network on the image segmented by the image segmentation unit 32 and transmits the detection result to the output unit 34 . For example, the analysis unit 33 may use an open source Automated M/L (DataRobot) deep learning algorithm. In addition, you can use H2O.ai, dotData, EdgeVerve, Aible, and Big Squid solutions.

Neural network learning selects a model through feature selection and algorithm selection from time-series data collected from input devices such as various sensors such as temperature, altitude, and fingerprints, cameras such as images and infrared rays, and LIDAR, and then repeats learning and performance verification processes. Repeat model selection through trial and error. After performance verification is complete, an artificial intelligence model is selected.

The analysis unit 33 performs a deep learning algorithm using a neural network to determine the sensor value, uses training data to learn the neural network, and verifies the performance of the neural network with test data.

The analysis unit 33 is suitable for object detection because it needs to find defects in the image and display the location, but deep learning classification technology is used while minimizing noise and pre-processing technology that divides the image in advance to effectively express the defect. Build the applied model.

The analysis unit 33 takes about 1 minute of pre-processing, 1 minute of detection, and 1 minute of output to detect defects in an image captured in an indoor space within 4 m, and can detect defects up to a size of 1 to 5 mm.

The output unit 34 transfers the defect probability value, which is the detection result of the analysis unit 33, to the interface unit 20 so that the numerical output of the interface unit 20 and the defect check procedure are performed.

The analysis unit 33 uses linear discriminant analysis (LDA), logistic regression (LR), support vector machine (SVM), decision tree (DT), k-nearest neighbor (KNN), random forest (RF), etc. Six machine learning algorithms can be applied.

Linear discriminant analysis (LDA) makes predictions by estimating the probability that a new set of inputs belongs to each class. The class that gets the highest probability is the output class and predictions are made. Logistic regression (LR) is the probability for two possible classification problems.

Support vector machines (SVMs) find the hyperplanar decision boundary that best splits the examples into two classes. The division is made smooth using margins where some points may be misclassified. A decision tree (DT) involves growing a tree to classify examples from a training data set. A tree can be thought of as dividing a training data set. Here, the example proceeds through the decision points of the tree until it reaches the leaves of the tree and is assigned a class label. A k-nearest neighbor (KNN) stores all available data and classifies new data points based on similarity. That is, when new data appears, it can be easily classified into well-sweet categories using k-Nearest Neighbors (KNN). Finally, a random forest (RF) consists of many decision trees for the final output, and is considered a very accurate and powerful method because the number of decision trees participating in the process is large.

Referring to FIG. 1 , the system authentication configuration includes a terminal 6 including a control unit 5 and an authentication server 7 .

The terminal 6 duplicates the data channels, receives the key value and biometric information of the terminal 6, and requests user authentication from the authentication server 7 through the first data channel. It is displayed on the display and transmitted to the authentication server (7).

The terminal 6 inputs the kit value displayed on the display of the terminal 6 and transmits it to the authentication server 7 through the second data channel along with user information. The terminal 6 requests the authentication server 7 to authenticate the system installed in the terminal 6 using the kit value and user information. The kit value of the terminal 6 can be generated from computer-specific information such as the CPU manufacturing number and the address of the Ethernet chip. The terminal 6 may obtain user information through face recognition using a camera, voice recognition using a microphone, and handwriting recognition using a display, and may use it for authentication.

The authentication server 7 receives the kit value from the terminal 6, receives the kit value and user information from the terminal 6 through a duplicated data channel, compares the kit value of the terminal 6 with the user information, and returns the user information Correspondingly, authentication for system use of the terminal 6 is processed. The authentication server 7 transmits the authentication result to the terminal 6 to allow the user to use the system. Due to the duplicated data channel of the terminal 6, loss of kit value can be minimized.

The authentication server 7 performs a history analysis of user information, and compares and determines consistency and change of user information over time. In the history analysis, if the user information shows consistency, the user's use is permitted, and if it shows change, the user's use is not permitted. When the user information shows consistency, the user's use of the system is permitted, thereby strengthening security to prevent a user whose user information has been altered from accessing the system.

The terminal 6, which is a means of authenticating the use of the system, does not directly connect to the system, but forms a detour through the authentication server 7, so that the network constituting the Internet network is composed of an internal network and an external network, and the IP address setting process When this is cumbersome, there is an advantage in that the authentication process using the terminal 6 is smoothly performed. At this time, the system is mounted on the terminal 6, the terminal 6 serves as an authentication terminal means, and the authentication server 7 serves as an authentication server means.

Figure 2 is a flow chart showing defect data collection and defect detection model construction used in the picture-based construction defect detection system using deep learning in Figure 1. The photo-based construction defect detection system using deep learning inputs a defect image Step (S101); Establishing an initial defect detection model for image segmentation and defect detection (S102); Determining defects in the divided images (S103); Secondarily classifying the image having a defect detection accuracy of 30 to 80% in step S103 (S104); Manually determining defects in the secondary classified images (S105); Collecting data of the image having a defect detection accuracy of 90% or more in step S103 and the image determined to be defective in step S105 (S106); monitoring the amount of learning of the defect detection model (S107); A step of securing deep learning learning data and building a defect detection model (S108) is performed.

3 is a flow chart showing a defect data reading test of the photo-based construction work defect detection system using deep learning in FIG. 1. Referring to FIG. starting the test (S201); Classifying the type of learning data (S202); Collecting defect-only images and foreground images (S203); Classifying the resolution (S204); Collecting 128, 245, 1024, 4096 size images (S205); Classifying the obstacle ratio (S206); Collecting images with obstacle ratios of 10%, 20%, 30%, and 40% (S207); Classifying defect types (S208); Collecting defect types A, B, C, and D (S209); is performed.

The device for analyzing defects in construction works checks the condition of the training image when only defects are trained or when an image containing the background is trained, and cross-verifies the defect reading accuracy while forcibly changing the image resolution to meet the minimum resolution condition of the image. , check the accuracy effect according to the ratio of surrounding obstacles based on the optimal resolution and image segmentation, and check whether there is a difference in optimal conditions for each defect type.

Figure 4 is an exemplary view showing a user interface of the picture-based construction defect detection system using deep learning in Figure 1. Referring to Figure 4, the user interface is selected AI defect inspection menu 101, the screen is switched to a picture Import (102) and history inquiry (103) are displayed, and when import picture (102) is selected, a list of images stored in the terminal (6) is displayed, and when an image (104) is selected from the image list, the screen is switched. Identification (105) and Reimport (106) are displayed, and when defect identification (105) is selected, defect identification (107) is displayed, a defect detection image (108) is displayed, and defect identification completion (109) is displayed. is displayed, and the screen is switched to display defect correct (110), non-defective (111), cancel (112), and when defect correct (110) is selected, defect identification completion (109) is displayed for the defect detection image (108) do.

5 is an exemplary view of defect data that can be used in the picture-based construction work defect detection system using deep learning in FIG. 1. Referring to FIG. 5, the photo-based construction work defect detection system using deep learning uses defect data Single defect images and defect data including obstacles can be used. When defect data including a single defect image or an obstacle is applied to a defect detection algorithm, a difference occurs in defect detection performance of the defect detection algorithm according to data characteristics, and the single defect image may be used to improve defect detection performance. In addition, if the obstacle ratio is low and the amount of data is large, defect data including obstacles may also be used.

A photo-based construction construction defect detection system using deep learning can cross-verify defect reading accuracy while forcibly changing the image resolution, and check the minimum resolution condition of the image to determine the resolution required for defect detection. setup is required. The photo-based construction defect detection system using deep learning has a resolution of 4032*3024 pixels, and can detect defects within 1 cm when shooting within a range of 4 m in width and looking straight at the surface to be filmed.

In the photo-based construction defect detection system using deep learning, it is possible to compensate by setting the accuracy threshold when there is no rapid change in accuracy even though it is affected by surrounding obstacles. Accuracy drops drastically.

A photo-based construction construction defect detection system using deep learning can recognize defects as defects when the image resolution for defect learning is too low, and can recognize shadowed areas as defects, so these images can be removed through pre-processing. .

The photo-based construction defect detection system using deep learning learns various wallpapers to improve defect detection performance, and learns normal images containing obstacles to further improve defect detection performance.

6 is an exemplary view showing image segmentation, defect detection, and detection results of the photo-based construction work defect detection system using deep learning in FIG. 1. Referring to FIG. 6, the photo-based construction work defect detection system using deep learning performs video segmentation processing on the indoor image of construction work, performs defect detection on the split image, and judges it as a defect when the defect probability is 56%, but can be supplemented through threshold correction, and recognizes the lower body when the defect probability is 94%. can be processed with success.

FIG. 7 is an exemplary diagram showing the detection results of image segmentation and defect detection applied to multiple indoor photos of the photo-based construction defect detection system using deep learning in FIG. 1 . Referring to FIG. 7 , photo-based using deep learning The building construction defect detection system displays a defect image 302 with a defect probability of 100% and a defect image 303 with a 48% probability for the building construction indoor image 301, and as another example, for the building construction indoor image 304 A defect image (305) with a defect probability of 99% is displayed, and as another example, a defect image (307) with a defect probability of 100%, a defect image (308) with a defect probability of 56%, and a defect image (308) with a defect probability of 69% are displayed for an indoor image (306) of construction work. A defect defect image 309 is displayed.

Figure 8 is a block diagram showing the configuration of the self-driving robot to which the picture-based building construction defect detection system using deep learning of Figure 1 is applied. Referring to Figure 8, the self-driving robot is photo-based using deep learning It includes the configuration of the construction work defect detection system and the autonomous driving unit 50, and communicates with the terminal 6.

In the self-driving unit 50 , the self-driving robot 40 autonomously travels in an indoor space for construction work, photographs the indoor space, and transmits the photographed image to the defect recognition unit 30 . The defect recognition unit 30 performs defect detection on the received image and transmits the detection result to the terminal 6 . The terminal 6 displays the defect detection result of the autonomous robot 40 on the screen.

The self-driving unit 50 localizes the position in the indoor space, measures the distance to the wall, and transmits an image captured when the self-driving robot 40 is located within a certain distance to the defect recognition unit 30, and the defect recognition unit (30) takes action to perform defect detection. In addition, the camera 10 captures an image by adjusting the focus of the lens and maximizes the performance of the imaging device.

Although the above has been described with reference to preferred embodiments of the present invention, those skilled in the art will variously modify and change the present invention within the scope not departing from the spirit and scope of the present invention described in the claims below. You will be able to fulfill what you can do.

1: processor 2: memory
3: input/output device 4: operating system
5: control unit 6: terminal
7: authentication server 10: camera
20: interface unit 30: defect recognition unit
31: still cut part 32: video division part
33: analysis unit 34: output unit
40: autonomous driving robot 50: autonomous driving unit

Claims (5)

User manipulation and user interface display display are processed, and user manipulations such as image selection and menu selection are transmitted to the defect recognition unit 30, and image segmentation, defect detection, and detection results, which are operating states of the defect recognition unit 30, are performed. an interface unit 20 displaying on a display; and
A defect recognition unit 30 that performs defect detection on an image of an indoor space where construction work is completed and includes a still cut unit 31, an image segmentation unit 32, an analysis unit 33, and an output unit 34. ); Photo-based construction defect detection system using deep learning, characterized in that it includes.
According to claim 1,
The defect recognition unit 30,
a still cut unit 31 that creates each still cut from the image selected in the interface unit 20 and transfers the still cut to the image segmentation unit 32 so that image segmentation and defect detection are performed on one still cut;
an image segmentation unit 32 that divides the still cut at a constant ratio horizontally and vertically, and sequentially transmits the divided images to the analysis unit 33 so that defects are detected on the divided images;
an analysis unit 33 which performs defect detection using a neural network on the image segmented by the image segmentation unit 32 and transmits a detection result to the output unit 34; and
An output unit 34 that transmits the defect probability value, which is the detection result of the analysis unit 33, to the interface unit 20 so that the interface unit 20 outputs a numerical value and performs a defect check procedure; Characterized by a photo-based construction defect detection system using deep learning.
According to claim 1,
The self-driving robot 40 further includes an autonomous driving unit 50 that autonomously travels in an indoor space for construction work, photographs the indoor space, and transmits the captured image to the defect recognition unit 30,
The defect recognition unit 30 performs defect detection on the received image, transmits the detection result to the terminal 6, and the terminal 6 receives the defect detection result of the autonomous robot 40. A photo-based construction defect detection system using deep learning, characterized in that it is displayed on the screen.
According to claim 1,
The analysis unit 33,
inputting a defect image (S101);
Establishing an initial defect detection model for image segmentation and defect detection (S102);
Determining defects in the divided images (S103);
Secondarily classifying the image having a defect detection accuracy of 30 to 80% in step S103 (S104);
Manually determining defects in the secondary classified images (S105);
Collecting data of the image having a defect detection accuracy of 90% or more in the step S103 and the image determined to be defective in the step S105 (S106);
monitoring the amount of learning of the defect detection model (S107); and
Securing deep learning learning data and building a defect detection model (S108); photo-based construction defect detection system using deep learning, characterized in that performing.
According to claim 1,
The analysis unit 33,
Starting a paper tear defect reading test (S201);
Classifying the type of learning data (S202);
Collecting defect-only images and foreground images (S203);
Classifying the resolution (S204);
Collecting 128, 245, 1024, 4096 size images (S205);
Classifying the obstacle ratio (S206);
Collecting images with obstacle ratios of 10%, 20%, 30%, and 40% (S207);
Classifying defect types (S208); and
A photo-based construction defect detection system using deep learning, characterized by performing the step (S209) of collecting defect types A, B, C, and D.

Images