KR20220161869A - Apparatus and method for remicon slump auto quality control using artificial intelligence - Google Patents

Abstract

A device for controlling a ready-mixed concrete slump automatic uniformity using artificial intelligence comprises: a feature extraction part that extracts a key feature to define a concentration of a ready-mixed concrete slump from a region-of-interest image, and configures a feature technical expert representing the key feature; a learning part that provides the learning data to a deep learning algorithm to determine how much the ready-mixed concrete has become according to the feature technical expert; and a recognition part that provides the feature technical expert of a real-time ready-mixed concrete image to the deep learning algorithm to derive an analysis result of the real-time ready-mixed concrete image. Therefore, the present invention is capable of reducing costs.

Description

Remicon slump automatic uniformity control device and method using artificial intelligence {APPARATUS AND METHOD FOR REMICON SLUMP AUTO QUALITY CONTROL USING ARTIFICIAL INTELLIGENCE}

The present invention relates to an apparatus and method for automatically uniformly controlling ready-mixed concrete slumps using artificial intelligence.

The slump is a value representing the consistency of ready-mixed concrete. When concrete is put in a cone with a height of 30cm and the cone is pulled out slowly, the apex goes down by its own weight according to the consistency of the dough.

In order to adjust the concentration (thinness) that determines the size of the ready-mixed concrete slump, add moisture, cement, sand, gravel, etc., and mix using a mixer or mixer. At this time, in order to check the concentration of ready-mixed concrete being manufactured in the mixer, a person must visually check the concentration of ready-mixed concrete inside the mixer. This method is highly likely to cause a problem of inconsistency in proper concentration due to human error in judgment, and a professional manpower must be on standby at all times to check the concentration. As a result, additional costs are required, and raw materials are wasted and lost due to momentary negligence and mistakes of experts, resulting in serious quality problems of the building.

The technical problem to be solved by the present invention is to provide a ready-mixed concrete slump automatic uniform control apparatus and method using artificial intelligence that can prevent waste and loss of raw materials due to human error and quality problems of resulting buildings.

An apparatus for automatically uniformly controlling ready-mixed concrete slump using artificial intelligence according to an embodiment of the present invention extracts key features for defining the concentration of ready-mixed concrete slump from an image of a region of interest, and extracts features constituting feature descriptors representing the key features. A learning unit that provides learning data to the deep learning algorithm to determine the degree of improvement of the ready-mixed concrete according to the feature descriptor, and a feature descriptor of the real-time ready-mixed corn image to the deep learning algorithm to determine the real-time ready-mixed concrete image It includes a recognition unit that derives analysis results.

The feature extraction unit may include a region-of-interest designation unit that separates the ready-mixed concrete image taken as a video frame by frame and extracts the region-of-interest image by specifying a region of interest of a certain size in the frame-based ready-mixed concrete image.

The feature extractor may further include a key feature extractor for transforming the ROI image into a frequency spectrum image through a Fourier transform.

The key feature extractor may select a feature image corresponding to one quadrant from the frequency spectrum image.

The feature extraction unit may further include a feature descriptor component configured to extract a frequency coefficient distribution as the feature descriptor by arranging the feature image in an order from the lowest frequency coefficient to the highest frequency coefficient.

The feature descriptor component may perform a convolution operation to remove outliers from the frequency coefficient distribution.

The feature extractor may further include a key feature extractor for extracting the ready-mixed concrete speed calculated for each frame from the ROI image as the key feature using a dense inverse search optical flow method.

The feature extraction unit may further include a feature descriptor component configured to structure a speed histogram formed by summing the ready-mixed concrete speeds calculated for each frame as the feature descriptor.

The feature extraction unit may further include a feature descriptor configuration unit for structuring a speed histogram formed by arranging the ready-mixed concrete speeds calculated for each frame in a row as the feature descriptor.

The feature extractor may further include a keypoint feature extractor extracting, as the keypoint feature, a difference image corresponding to a difference between a previous frame and a current frame of the ROI image.

The feature extraction unit may further include a feature descriptor component configured to structure the difference image as the feature descriptor.

The learning unit may provide the feature descriptor and label information corresponding to the feature descriptor as training data of the deep learning algorithm.

The analysis result may include label information corresponding to the feature descriptor of the real-time ready-mixed concrete image.

The ready-mixed concrete slump automatic uniformity control device using artificial intelligence may further include an automatic uniformity control component that automatically determines whether ready-mixed concrete material is added based on the analysis result of the real-time ready-mixed corn image and adds or subtracts the deficient material.

According to another embodiment of the present invention, a method for automatically uniformly controlling a ready-mixed concrete slump using artificial intelligence divides a ready-mixed concrete image captured as a video into frame units, and designates a region of interest of a certain size in the ready-mixed concrete image in each frame to obtain an image of the region of interest. Extracting, extracting key features for defining the concentration of the ready-mixed concrete slump from the region-of-interest image, constructing feature descriptors representing the key features, determining how much the ready-mixed concrete has become according to the feature descriptor Providing learning data to the deep learning algorithm so as to be able to do so, and providing feature descriptors of the real-time ready-mixed corn image to the deep learning algorithm to derive an analysis result of the real-time ready-mixed corn image.

The extracting of the key feature may include transforming the ROI image into a frequency spectrum image through a Fourier transform, and selecting a feature image corresponding to one quadrant from the frequency spectrum image.

The configuring of the feature descriptor may include extracting a frequency coefficient distribution as the feature descriptor by arranging the feature descriptor in an order from the lowest frequency coefficient to the highest frequency coefficient in the feature image.

Configuring the feature descriptor may further include performing a convolution operation to remove outliers from the frequency coefficient distribution.

The step of extracting the keypoint features may include extracting, as the keypoint features, a ready-mixed concrete speed calculated for each frame in the ROI image by a dense reverse search optical flow method.

Configuring the feature descriptor may include structuring a speed histogram formed by summing or arranging the ready-mixed concrete speeds calculated for each frame as the feature descriptor.

The extracting of the keypoint feature may include extracting a difference image corresponding to a difference between a previous frame and a current frame of the ROI image as the keypoint feature.

Providing the learning data to the deep learning algorithm may include providing the feature descriptor and label information corresponding to the feature descriptor as training data of the deep learning algorithm.

The analysis result may include label information corresponding to the feature descriptor of the real-time ready-mixed concrete image.

The automatic uniform control method for ready-mixed concrete slump using artificial intelligence may further include automatically determining whether ready-mixed concrete material is added based on the analysis result of the real-time ready-mixed corn image and adding or subtracting the missing material.

An apparatus and method for automatically uniformly controlling ready-mixed concrete slump using artificial intelligence according to an embodiment of the present invention receives an image inside the mixer in real time to determine the concentration of the required slump and at the same time determine the deficient material to command the addition or subtraction of the deficient material can

Unlike the existing method in which an expert had to stand by at all times and directly check the concentration of ready-mixed concrete, the artificial intelligence-based automatic uniform control device and method for ready-mixed corn slump enables production of ready-mixed concrete at an appropriate concentration through direct judgment and proper concentration The probability of inconsistency can be minimized and costs can be reduced.

Remicon slump automatic uniformity control device and method using artificial intelligence extracts key point features in image data and uses a criterion generation mechanism through machine learning to classify minute differences in concentration that cannot be classified by the human eye. It can be used not only for ready-mixed concrete but also for determining the concentration of various liquid raw materials, and its technology and system scalability and its applicability are high.

With the direct management of the artificial intelligence recognition system, it is possible to identify the clear cause of errors that may occur in the process of identifying the mixing ratio, inputting materials and water, and operating the mixer, and promptly responding to problems that may occur in the manufacturing process. At the same time, the loss that may occur in the manufacturing process can be minimized by prompt restart.

The artificial intelligence recognition system recognizes and classifies quality by itself, so quality control is possible without professional knowledge or technology, and a small number of system administrators who have received simple training can manage and operate the entire production process. Therefore, the ready-mixed concrete manufacturing process can be performed with a small number of manpower.

The artificial intelligence recognition system provides a fully automated alternative to the previously maintained partial process automation, and through complete process automation, one of the elements required by smart factories, the effects of cost reduction, production time reduction, and maintenance and improvement of production quality are achieved. can provide Through this, it is possible to reach the ranks of the ready-mixed concrete smart plant operation management technology and system that are attracting attention worldwide.

1 is a block diagram showing an apparatus and method for automatically uniformly controlling a ready-mixed concrete slump using artificial intelligence according to an embodiment of the present invention.
2 is an exemplary diagram illustrating a process of designating a region of interest according to an embodiment of the present invention.
3 is an exemplary diagram illustrating a process of Fourier transforming a region of interest according to an embodiment of the present invention.
4 is an exemplary diagram illustrating a process of selecting a key feature image from a frequency image according to an embodiment of the present invention.
5 is an exemplary diagram illustrating a process of constructing a feature descriptor according to an embodiment of the present invention.
6 is an exemplary diagram illustrating a process of removing outliers from a feature descriptor according to an embodiment of the present invention.
7 is an exemplary diagram illustrating a process of providing a feature descriptor as training data of a deep learning classifier of a neural network according to an embodiment of the present invention.

Hereinafter, with reference to the accompanying drawings, embodiments of the present invention will be described in detail so that those skilled in the art can easily carry out the present invention. This invention may be embodied in many different forms and is not limited to the embodiments set forth herein.

In order to clearly describe the present invention, parts irrelevant to the description are omitted, and the same reference numerals are assigned to the same or similar components throughout the specification.

In addition, throughout the specification, when a certain component is said to "include", it means that it may further include other components without excluding other components unless otherwise stated.

Hereinafter, a ready-mixed concrete slump automatic uniform control apparatus and method using artificial intelligence according to an embodiment of the present invention will be described with reference to FIGS. 1 to 7.

1 is a block diagram showing an apparatus and method for automatically uniformly controlling a ready-mixed concrete slump using artificial intelligence according to an embodiment of the present invention. 2 is an exemplary diagram illustrating a process of designating a region of interest according to an embodiment of the present invention. 3 is an exemplary diagram illustrating a process of Fourier transforming a region of interest according to an embodiment of the present invention. 4 is an exemplary diagram illustrating a process of selecting a key feature image from a frequency image according to an embodiment of the present invention. 5 is an exemplary diagram illustrating a process of constructing a feature descriptor according to an embodiment of the present invention. 6 is an exemplary diagram illustrating a process of removing outliers from a feature descriptor according to an embodiment of the present invention. 7 is an exemplary diagram illustrating a process of providing a feature descriptor as training data of a deep learning classifier of a neural network according to an embodiment of the present invention.

1 to 7, the automatic uniform control apparatus for ready-mixed corn slump using artificial intelligence according to an embodiment of the present invention collects ready-mixed concrete mixture image data for each concentration inside the mixer, defines and extracts image characteristics for each concentration, and performs machine learning A concentration determination model may be trained based on an algorithm, for example, a deep learning algorithm. The system learned in this way can recognize the concentration of ready-mixed concrete, and then the ready-mixed concrete slump automatic uniformity control device receives the ready-mixed concrete image inside the mixer in real time to determine the concentration of the slump and at the same time determine the deficiency material and automatically command the addition or subtraction of the missing material. can

In more detail, the ready-mixed concrete slump automatic uniformity control device using artificial intelligence may include a feature extraction unit 100, a learning unit 200, a recognition unit 300, and an automatic uniformity control component 400.

The feature extraction unit 100 extracts key features for the ready-mixed concrete image and performs a process of constructing feature descriptors representing the key features. The feature extraction unit 100 may include an input image collection unit 101 , a region of interest designation unit 102 , a key feature extraction unit 103 and a feature descriptor configuration unit 104 .

The input image collection unit 101 may transmit the ready-mixed concrete image 10 using a camera to the region of interest designation unit 102 . The input image collection unit 101 may take a video of the ready-mixed concrete image 10 periodically or for a certain period of time. The ready-mixed concrete image (input image) 10 collected by the input image collection unit 101 is for constructing key features and feature descriptors.

The region of interest designation unit 102 separates the remicon image 10 in the process of mixing the remicon into frame units, and designates a region of interest 11 of a certain size from the remicon image 10 in frame units to obtain an image of the region of interest ( 12) can be extracted. The region of interest 11 may be a portion where feature points are visible from among the entire image of the ready-mixed concrete image 10. The region of interest 11 may be designated with a constant size based on the central portion of the ready-mixed concrete image 10 in units of frames. The region of interest designator 102 does not designate the region of interest 11 in all frames of the ready-mixed concrete image 10, acquires frames from the ready-mixed concrete image 10 at regular intervals, and obtains a frame-by-frame ready-mixed concrete image 10 A region of interest 11 can be designated in This is because even if a portion is extracted from the ready-mixed-corn image 10, it can represent the entire ready-mixed-corn image 10, and this is to facilitate real-time processing by reducing the number of unnecessary features. The ROI designator 102 transfers the ROI image 12 composed of the ROI to the keypoint feature extraction unit 103 .

The key feature extractor 103 may extract key features (pattern definition) for defining the concentration of the ready-mixed concrete slump from the ROI image 12 . The keypoint feature extractor 103 may define keypoint features as a frequency configuration distribution of the ROI image 12 .

Specifically, the key feature extractor 103 may transform the ROI image 12 into a frequency spectrum image 13 through Fourier transform. The frequency spectrum image 13 may include a high-frequency region in which pixel brightness changes rapidly and a low-frequency region in which pixel brightness changes little or gradually. The Fourier transform is an algorithm that transforms the ROI image 12 (original image) in units of frames into the frequency spectrum image 13.

Also, the key feature extractor 103 may select a feature image 14 corresponding to one quadrant from the frequency spectrum image 13 . Since the frequency spectrum image 13 has a symmetrical shape in the second quadrant, the third quadrant, and the fourth quadrant based on the first quadrant, all information of the original image can be expressed in one quadrant among the four quadrants. 4 illustrates that the second quadrant is selected in the frequency spectrum image 13, and in the feature image 14 of the second quadrant, the lowest frequency corresponds to the upper left and the highest frequency corresponds to the lower right.

Depending on the embodiment, the keypoint feature extractor 103 may extract keypoint features using various methods other than Fourier transform. For example, the key feature extractor 103 includes Variance of Variance Components (VVC), Variance of Variance (VOV), Scale-Invariant Feature Transform Feature (SIFT), Speed-Up Robust Features (SURF), and Histogram of HOG (Histogram of Variance). Oriented Gradient), FAST(Features from Accelerated Segment Test), BRIEF(Binary Robust Independent Elementary Features), ORB(Oriented FAST and Rotated BRIEF), Haar Feature, Ferns Feature, MCT(Modified Census Transform) Feature, GLCM(Gray Level Co Key features can be extracted using key feature extraction algorithms such as -occurrence Matrix (Local Binary Pattern), LBP (Local Binary Pattern), and Gabor Filter.

VVC uses variance to indicate whether a sample group extracted from a population reflects the characteristics of the population well. Variance is a feature that uses the sum and average to represent the characteristics of a group, so it represents the characteristics of the group well. Therefore, if the subgroup reflects the characteristics of the population, the variances of the two will have similar values. VVC determines similarity by comparing the variance of subgroups with the variance of the population.

VOV is a method that improves the characteristics of VVC, and sets the population as the background of the image and the sample group as the image containing defects, and compares their variance factors to determine the presence or absence of defects. Since the brightness value of the sample group including the defect in the image shows contrast with the brightness value of the sample group including only the background, the existence and location of the defect are determined by comparing these two subgroups.

SIFT is an algorithm that extracts corner feature points regardless of the size of the image. SIFT enlarges and reduces the original image to construct four sets of images (Octave), and detects edges in the image through DoG (Difference of Gaussian) operation using several Gaussian masks with different standard deviation values targeting them. After that, based on the edge point with the largest DoG value among the adjacent edge points, the edge point with the largest horizontal and vertical differential values is determined as a corner point.

SURF is an algorithm that constructs an integral image and detects candidate corner points using a Hessian matrix using second order derivative. SURF quickly computes Hessian matrix operations by applying an approximated filter to an integral image without using a general Gaussian filter. In addition, corner points of various sizes are detected by changing the size of this approximated filter. SURF can increase the calculation speed by changing the size of the integral image, approximated Hessian matrix, and filter.

HOG is a method of obtaining a gradient direction of an image by considering the change in angle and size of a gray scale for a pixel, and extracting feature values histogrammed from the gradient direction of the image as a frequency for each angle. First, a Sobel mask is used to detect edge points through differentiation in an image. A histogram is formed by grouping the gradient directions of the obtained image in units of blocks.

FAST is an algorithm that quickly detects corner feature points in an image. For all pixel points in the image, if there is a large difference in brightness between pixels 1, 5, 9, and 13 among 16 neighboring pixels and the center pixel, that point is designated as a corner feature point candidate. Corner feature points in the image are determined by constructing a decision tree ID3 for corner feature point candidates detected from several images.

BRIEF is an algorithm that constructs a binary descriptor by comparing the brightness of an arbitrary pair of pixels within a patch. It has the advantage of fast calculation, high discrimination, and fast matching process, but has the disadvantage of being vulnerable to rotation. It is not an algorithm that detects feature points such as corners and edges in an image, so it must be used together with an algorithm that detects feature points.

ORB is an algorithm that constructs feature descriptors by applying the FAST algorithm that applies image pyramids to detect corner points for images of various sizes, and defining BRIEF descriptors supplemented by rotation transformation of a matrix based on corner points. Complete corner points are detected by applying the Harris corner algorithm based on the candidate points through the FAST algorithm.

Haar feature is a feature using the brightness difference between regions in an image. Each feature value is calculated by subtracting the sum of the brightness of the black part from the sum of the brightness of the image pixels corresponding to the white part. An object is identified by combining a plurality of features extracted in this way. It is less sensitive to changes in the shape and position of objects within the area, but is sensitive to changes in image rotation, contrast, and brightness.

The Ferns Feature first selects feature points from the image, obtains a local patch centered on each feature point, then grabs a pair of two points at random within the local patch and determines whether the pixel brightness difference between the two points is negative or positive. . It is robust to image contrast and brightness changes, but is sensitive to changes in shape or rotation.

MCT is used for converting an input image into an image from which the influence of ambient brightness change is removed. The MCT feature descriptor for a pixel is a value obtained by encoding 0 if the brightness of a pixel in the surrounding area is brighter than the average of the local area and 1 if it is darker, and connecting the result to a bit string, and is matched by the hamming distance through XOR. . It is characterized by being robust against changes in brightness.

The GLCM matrix is a method of expressing an image as a histogram of brightness values by using the brightness values of the image as coordinates. Values of the current pixel and neighboring pixels in the image are used as x, y coordinates, and values are added to the GLCM matrix of the corresponding coordinates. Therefore, when the background is expressed in a similar color, the histogram for the background color is expressed as a high value, but when a specific value such as a defect is input, it is recognized as an error and a defect can be detected.

LBP is a feature of an efficient image that is used for expressing the texture of an image. The LBP operator computes local binary pattern values. After calculating LBP values for all pixels, a normalized histogram of LBP values is created. The texture of the image is defined as a histogram consisting of 256 bins, and the defined texture is used to detect defects by the histogram matching method.

A Gabor filter is a filter that combines a cosine function and a Gaussian function. The shape of the filter can be effectively configured using the five setting values of λ, θ, ψ, σ, and γ considering the brightness of the image, the frequency and shape of the pattern, and the direction. It is excellent in expressing the size and direction of surface texture and shows high performance in backgrounds with complex patterns, so it is used in defect inspection.

On the other hand, the key feature extractor 103 applies a dense inverse search (DIS) optical flow method to the region of interest image 12 to extract key features for defining the concentration of the ready-mixed concrete slump. can If the concentration (viscosity) of the slump is high, the mixing speed of the ready-mixed concrete (remicon speed) slows down, and if the concentration (viscosity) of the slump is low, the speed of ready-mixed concrete increases. By doing so, the concentration of the slump can be extracted. The DIS optical flow is to calculate the ready-mixed concrete velocity in the region of interest image 12 after developing it through a Taylor series after assuming that the brightness values of feature points and the brightness values of neighboring pixels of the region of interest image 12 do not change. That is, the key point feature extractor 103 may extract the ready-mixed concrete speed calculated for each frame from the ROI image 12 as a key point feature using the DIS optical flow method.

On the other hand, the keypoint feature extraction unit 103 may compare the previous frame and the current frame of the ROI image 12 to generate a difference image corresponding to the difference, and extract the difference image as a keypoint feature. The difference image may display the waveform and flow of ready-mixed concrete.

The feature descriptor constructing unit 104 receives the key feature from the key feature extraction unit 103 and constructs a feature descriptor representing the key feature. More specifically, the feature descriptor component 104 receives the feature image 14 of the frequency spectrum image 13 converted by the key feature extractor 103, and the lowest frequency coefficient in the feature image 14. It is possible to extract the frequency coefficient distribution by sorting in order of the highest frequency coefficient from . As illustrated in FIG. 5 , frequency coefficients in the feature image 14 are arranged in a zigzag order in a diagonal direction. To sort from the lowest to the highest frequency coefficient, the lowest frequency coefficient is located at (0,0), so the frequency coefficients are (0,0), (0,1), (1,0), (2,0) , (1,1), (0,2), ... can be sorted in the order.

Since the distribution degree and speed of the brightness change of the ready-mixed concrete images vary according to the thickness of the ready-mixed concrete, each ready-mixed concrete image having a different thickness has a frequency coefficient distribution that can be distinguished from each other. Therefore, the frequency coefficient distribution can be referred to as a representative value representing the characteristics of the ready-mixed concrete image.

The feature descriptor component 104 may structure the frequency coefficient distribution as a feature descriptor (FD) representing the ready-mixed concrete. The feature descriptor component 104 transmits the feature descriptor (FD) to the learning unit 200 to be used as learning data of the deep learning classifier, and transmits the feature descriptor (FD) to the recognition unit 300 to make the ready-mixed image ( 10) can be analyzed in real time.

On the other hand, there are times when an outlier in which a part of a value jumps in the frequency coefficient distribution is identified. Therefore, as illustrated in FIG. 6, the feature descriptor component 104 configures a filter to effectively remove outliers without making all data characteristics of the feature descriptor (FD) disappear and performs convolution operation. can do. The feature descriptor constructing unit 104 may construct a data set with convolutionally operated feature descriptors. These data sets can be used for deep learning learning to further improve classification accuracy.

On the other hand, the feature descriptor composition unit 104 is structured as a feature descriptor (FD) with a speed histogram formed by adding the ready-mixed concrete speeds calculated for each frame in the region of interest image 12 by the DIS optical flow method in the key feature extractor 103 can do.

Alternatively, the feature descriptor component 104 is a feature descriptor (FD) of the speed histogram formed by arranging the ready-mixed concrete speeds calculated for each frame in the region of interest image 12 by the DIS optical flow method in the key feature extractor 103. can be structured as

On the other hand, the feature descriptor configuration unit 104 may structure a difference image corresponding to a difference between a previous frame and a current frame of the ROI image 12 as a feature descriptor (FD).

The learning unit 200 provides learning data to the neural network so that it can determine how much the ready-mixed concrete has become according to the feature descriptor (FD) using a deep learning algorithm, and the degree of becoming in the ready-mixed corn image 10 according to the learning result. Performs the process of confirming the quantitative standard for discriminating. The learning data is at least one of a frequency coefficient distribution, a speed histogram formed by adding up the ready-mixed concrete speeds calculated for each frame, a speed histogram formed by listing the ready-mixed concrete speeds calculated for each frame in a row, and a difference image corresponding to the difference between the previous frame and the current frame. It may be made based on a feature descriptor (FD) including.

The learning unit 200 may include a deep learning classifier learning unit 201 and a quantitative standard determining unit 202 .

The deep learning classifier learning unit 201 provides feature values (feature descriptors) of a classification target required for learning by a deep learning algorithm (or deep learning classifier) of a neural network and label information corresponding to the feature values. For example, the deep learning classifier learning unit 201 is a feature descriptor (FD) and a first label (eg, label 0) for the ready-mixed corn image 10 having the largest becoming, and the ready-mixed corn image 10 having an intermediate becoming. ) for the feature descriptor (FD) and the second label (e.g., label 1), the feature descriptor (FD) and the third label (e.g., label 2) for the dilute ready-mixed concrete image (10) having the smallest probability can be provided as training data of the deep learning algorithm.

As illustrated in FIG. 7, the deep learning algorithm encodes the provided label information in a one-hot format and provides it to a deep learning learning algorithm, PackPropagation Learning Algorithm. For example, the vector [1,0,0] for label 0, the vector [0,1,0] for label 1, and the vector [0,0,1] for label 2 are transformed into the training data. can be used as a classification criterion.

The deep learning algorithm that has been trained has learned weights that are different from the initially set weights. The quantitative standard determination unit 202 may determine the learned weight as a mortar quantitative standard for automatically determining the degree of becoming of the ready-mixed concrete image 10 entering as a real-time input.

The recognizing unit 300 recognizes the feature descriptor (FD) of the ready-mixed-corn image 10 input in real time based on quantitative criteria and performs a process of analyzing the becoming of ready-mixed concrete. Recognition unit 300 may include an analysis unit 301 for receiving the feature descriptor (FD) of the remicon image 10 from the feature descriptor component 104 and analyzing the becoming of the remicon.

The analysis unit 301 may provide the feature descriptor (FD) of the ready-mixed concrete image 10 input in real time to the deep learning algorithm, or provide the determined quantitative criteria together with the feature descriptor (FD) to the deep learning algorithm. . The deep learning algorithm may analyze the remicon based on the determined quantitative standard and provide the analysis result to the analysis unit 301. In this case, since the deep learning algorithm is not learning, the analyzer 301 does not need to provide label information for the input feature descriptor (FD). And since the error backpropagation learning algorithm is not applied, the deep learning algorithm can derive analysis results at a very high speed. The analysis result includes label information corresponding to the feature descriptor (FD) of the ready-mixed concrete image 10 input in real time, and this label information may be label information provided along with the feature descriptor (FD) during learning.

For example, if the characteristic descriptor (FD) of the ready-mixed-corn image 10 input in real time corresponds to the feature descriptor (FD) of the ready-mixed-corn image 10 with the greatest value, the ready-mixed-corn image with the greatest value during learning ( 10) may be derived as an analysis result of the provided label information (eg, label 0) along with the feature descriptor (FD).

The analysis unit 301 may transmit the analysis result of the ready-mixed concrete image 10 input in real time to the automatic uniform control component 400. Alternatively, the analysis unit 301 may inform the user of the analysis result of the ready-mixed concrete image 10 input in real time through a display.

The automatic uniformity control component (Auto Quality Control) 400 automatically determines whether to add ready-mixed concrete materials such as water, cement, sand, gravel, etc. based on the analysis result of the ready-mixed concrete image 10, and adds or subtracts deficient materials to ready-mixed concrete process can be performed. The automatic uniformity control component 400 may include an automatic uniformity control unit 401 .

The automatic uniform control unit 401 may determine the state of the ready-mixed concrete currently being manufactured based on the analysis result of the ready-mixed concrete image 10, and determine whether to add the ready-mixed concrete material. When the ready-mixed concrete is completed, the automatic uniform control unit 401 may inform the user of the status of the ready-mixed concrete through a display or the like so that the ready-mixed concrete can be placed.

As described above, the apparatus and method for automatically uniformly controlling ready-mixed concrete slump using artificial intelligence according to an embodiment of the present invention receives an image inside the mixer in real time to determine the concentration of the required slump and at the same time determine the deficient material to determine the deficient material. You can command the addition or subtraction of

The drawings and detailed description of the present invention referred to so far are only examples of the present invention, which are only used for the purpose of explaining the present invention, and are used to limit the scope of the present invention described in the meaning or claims. It is not. Therefore, those skilled in the art will understand that various modifications and equivalent other embodiments are possible therefrom. Therefore, the true technical protection scope of the present invention should be determined by the technical spirit of the appended claims.

100: feature extraction unit 101: input image collection unit
102: region of interest designation unit 103: key feature extraction unit
104: feature descriptor component 200: learning unit
201: deep learning classifier learning unit 202: quantitative standard determination unit
300: recognition unit 301: analysis unit
400: automatic uniformity control unit 401: automatic uniformity control unit

Claims (24)

a feature extraction unit that extracts key features for defining the concentration of the ready-mixed concrete slump from the region-of-interest image and constructs a feature descriptor representing the key features;
A learning unit for providing learning data to a deep learning algorithm to determine how much the ready-mixed concrete has become according to the feature descriptor; and
Remicon slump automatic uniform control device using artificial intelligence including a recognition unit for providing the feature descriptor of the real-time ready-mixed-corn image to the deep learning algorithm to derive the analysis result of the real-time ready-mixed-corn image. According to claim 1,
The feature extraction unit divides the ready-mixed concrete image taken as a video frame by frame, and designates a region of interest of a certain size from the ready-mixed concrete image in frame units to extract the region of interest image Remicon slump using artificial intelligence including a region of interest designation unit Automatic uniformity control device. According to claim 2,
The feature extraction unit further comprises a key feature extraction unit for converting the image of the region of interest into a frequency spectrum image by Fourier transform. According to claim 3,
The key feature extraction unit automatically uniformly controls the ready-mixed concrete slump using artificial intelligence for selecting a feature image corresponding to one quadrant in the frequency spectrum image. According to claim 4,
The feature extraction unit further comprises a feature descriptor component for extracting a frequency coefficient distribution as the feature descriptor by arranging it in order from the lowest frequency coefficient to the highest frequency coefficient in the feature image. According to claim 5,
The feature descriptor component unit automatically uniform control device for ready-mixed concrete slump using artificial intelligence that performs convolution operation to remove outliers from the frequency coefficient distribution. According to claim 2,
The feature extraction unit further comprises a key feature extraction unit for extracting the ready-mixed concrete speed calculated for each frame from the region-of-interest image as the key point feature by a dense reverse search optical flow method. According to claim 7,
The feature extraction unit further comprises a feature descriptor configuration unit for structuring a speed histogram formed by summing the remicon speeds calculated for each frame as the feature descriptor. According to claim 7,
The feature extraction unit further comprises a feature descriptor configuration unit for structuring a speed histogram formed by arranging the remicon speeds calculated for each frame in a row as the feature descriptor. According to claim 2,
The feature extraction unit further comprises a key point feature extraction unit for extracting a difference image corresponding to a difference between a previous frame and a current frame of the region of interest image as the key point feature. According to claim 10,
The feature extraction unit further comprises a feature descriptor component for structuring the difference image as the feature descriptor. According to claim 1,
The learning unit provides the feature descriptor and label information corresponding to the feature descriptor as learning data of the deep learning algorithm. According to claim 12,
The analysis result is a ready-mixed concrete slump automatic uniform control device using artificial intelligence containing label information corresponding to the feature descriptor of the real-time ready-mixed concrete image. According to claim 1,
Based on the analysis result of the real-time ready-mixed-corn image, the ready-mixed concrete slump automatic uniform control device using artificial intelligence further comprising an automatic uniformity control component for automatically determining whether to add ready-mixed concrete material and adding or subtracting deficient materials. Separating the ready-mixed concrete image taken as a video frame by frame, and extracting the region-of-interest image by designating a region of interest of a certain size in the frame-based ready-mixed concrete image;
Extracting key features for defining the concentration of the ready-mixed concrete slump from the region-of-interest image;
constructing a feature descriptor representing the key feature;
Providing learning data to a deep learning algorithm to determine how much the ready-mixed concrete has become according to the feature descriptor; and
An automatic uniform control method for ready-mixed concrete slumps using artificial intelligence, comprising the step of providing a feature descriptor of the real-time ready-mixed-corn image to the deep learning algorithm to derive an analysis result of the real-time ready-mixed-corn image. According to claim 15,
The step of extracting the key features is,
converting the ROI image into a frequency spectrum image by Fourier transform; and
Automatic uniform control method for ready-mixed concrete slump using artificial intelligence comprising the step of selecting a feature image corresponding to one quadrant from the frequency spectrum image. According to claim 16,
The step of constructing the feature descriptor,
Remicon slump automatic uniform control method using artificial intelligence comprising the step of extracting a frequency coefficient distribution as the feature descriptor by arranging it in order from the lowest frequency coefficient to the highest frequency coefficient in the feature image. According to claim 17,
The step of constructing the feature descriptor,
Automatic uniform control method for ready-mixed concrete slump using artificial intelligence, further comprising performing a convolution operation to remove outliers from the frequency coefficient distribution. According to claim 15,
The step of extracting the key features is,
A method for automatically uniformly controlling ready-mixed concrete slumps using artificial intelligence, comprising the step of extracting the ready-mixed concrete speed calculated for each frame in the region of interest image by a dense reverse search optical flow method as the key feature. According to claim 19,
The step of constructing the feature descriptor,
Remicon slump automatic uniform control method using artificial intelligence comprising the step of structuring a speed histogram formed by summing or lining up the ready-mixed concrete speeds calculated for each frame as the feature descriptor. According to claim 15,
The step of extracting the key features is,
Automatic uniform control method for ready-mixed concrete slump using artificial intelligence, comprising the step of extracting a difference image corresponding to the difference between the previous frame and the current frame of the region of interest image as the key feature. According to claim 15,
Providing the learning data to the deep learning algorithm,
A method for automatically uniformly controlling ready-mixed concrete slump using artificial intelligence comprising the step of providing the feature descriptor and label information corresponding to the feature descriptor as learning data of the deep learning algorithm. 23. The method of claim 22,
The analysis result is an automatic uniform control method for ready-mixed concrete slump using artificial intelligence including label information corresponding to the feature descriptor of the real-time ready-mixed concrete image. According to claim 15,
Based on the analysis result of the real-time ready-mixed-corn image, automatically determining whether to add ready-mixed-corn materials and adding or subtracting deficient materials.

Images