KR102079378B1 - Method and apparatus for reconstructing of video - Google Patents

Abstract

An image restoration method and apparatus are provided. The image reconstruction method extracts a predetermined type of image characteristic for each frame of an input image file, generates a characteristic vector, maps the characteristic vector to a low dimensional space through dimensional reduction, and distributes the characteristic vectors according to the visualization. Generate a model graph optimized for the RX, reconstruct the order of the frames based on the model graph and the characteristic vector for each frame, and restore the image file based on the reconstructed order of the frames.

Description

Image restoration method and apparatus therefor {METHOD AND APPARATUS FOR RECONSTRUCTING OF VIDEO}

The present invention relates to a method and apparatus for restoring an image, and more particularly, to a method and apparatus for restoring a sequence of video frames through visual analysis.

As vehicle black boxes are becoming more popular, images of black boxes are being used as a major evidence in the judicial and insurance industries in case of traffic accidents. However, automotive black box devices may not always provide evidence video intact. For example, when the power supply to the vehicle black box device is temporarily cut off, the image encoding may fail and the image may not be recorded intact. In addition, black box image data is frequently lost due to random deletion of a user. In order for the damaged black box image to be effective as evidence, it must be restored to an intact image file of visually identifiable level through separate work.

On the other hand, as the most popular image file recovery technique, there is a file carving technique that restores a file by searching for a specific pattern using metadata of the file. However, large image files are often fragmented in the middle, so it is difficult to completely restore them by the file carving technique. In order to solve this problem, various image file recovery techniques have been developed, such as a technique for restoring based on codec signature, but there is still a difficulty in restoring file by file due to the absence or partial omission of image order information.

Korean Patent Registration No. 10-1348216 (Invention name: Vehicle image recording device and image recording method)

An embodiment of the present invention is to provide a method and apparatus for restoring the order of fragmented image data by reconstructing the damaged image but visually analyzing the interrelation between each frame based on the restored image frame.

However, the technical problem to be achieved by the present embodiment is not limited to the technical problem as described above, and other technical problems may exist.

As a technical means for achieving the above-described technical problem, an image restoration method according to the first aspect of the present invention, the step of extracting a predetermined type of image characteristics for each frame of the input image file to generate a feature vector; Visualizing the feature vector by mapping it to a low dimensional space through a dimension reduction process; Generating a model graph optimized for distribution of the feature vectors according to the visualization, and reconstructing the order of the frames in the image based on the model graph and the feature vector for each frame; And restoring an image file based on the order of the reconstructed frames.

In addition, the image restoration apparatus according to the second aspect of the present invention, the memory for storing the image restoration program; And a processor configured to execute a program stored in the memory, wherein the processor extracts a predetermined type of image characteristic for each frame of an input image file according to the execution of the image restoration program to generate a characteristic vector. Is visualized by mapping to a low dimensional space through a dimensional reduction process, generating a model graph optimized for distribution of feature vectors according to the visualization, and ordering the images in the frames based on the model graph and the feature vector for each frame. And reconstruct the image file based on the reconstructed frames.

A third aspect of the invention also provides a computer readable recording medium having recorded thereon a program for implementing the method of the first aspect.

According to any one of the above-described means for solving the problem of the present invention, when any image is fragmented in the frame unit in the storage device, etc., the result of automatically recombining the image file unit without encoding information or metadata of the corresponding image is provided. can do.

In addition, automated frame recombination may be performed regardless of the number of frames included in the damaged image, and thus may be effectively used for securing evidence in the digital forensic region.

1 is a block diagram of an image restoration apparatus according to an embodiment of the present invention.
2 is an example for describing a color characteristic and a geometric characteristic extraction method for each image frame block according to an embodiment of the present invention.
3 is an example for describing a vanishing point-based block generation method according to an embodiment of the present invention.
4 is an example of a result of visualizing a feature vector of an image frame according to an embodiment of the present invention.
5 is an example of a result of visualizing a two-dimensional feature vector according to an embodiment of the present invention through a curve-fitting technique.
6 is a flowchart illustrating an image restoration method according to an embodiment of the present invention.

DETAILED DESCRIPTION Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those skilled in the art may easily implement the present invention. As those skilled in the art would realize, the described embodiments may be modified in various different ways, all without departing from the spirit or scope of the present invention. In the drawings, parts irrelevant to the description are omitted in the drawings, and like reference numerals designate like parts throughout the specification. In addition, while describing with reference to the drawings, even if the configuration shown by the same name may be different according to the drawing number, the drawing number is just described for convenience of description and the concept, features, functions of each configuration by the corresponding reference number Or the effects are not to be construed as limiting.

Throughout the specification, when a portion is said to "include" any component, it means that it can further include other components, without excluding other components, unless specifically stated otherwise, one or more It is to be understood that the present invention does not exclude in advance the possibility of the presence or the addition of other features, numbers, steps, operations, components, parts, or combinations thereof.

In the present specification, the term 'part' or 'module' includes a unit realized by hardware or software and a unit realized by both, and one unit is realized by using two or more pieces of hardware. Two or more units may be implemented by one hardware.

1 is a block diagram of an image restoration apparatus according to an embodiment of the present invention.

As shown in FIG. 1, the image restoration apparatus 100 includes an input module 110, a memory 120, a processor 130, and an output module 120.

The input module 110 receives an image file, which is a restoration process target, and delivers the image file to the processor 130.

In this case, the input module 110 may include a function of a communication module for transmitting and receiving data by performing communication with an external device (for example, a vehicle black box device, etc.) interlocked in advance. The image file input through the input module 110 is an image file photographed or generated through an external device and may include damaged image data.

The memory 120 stores various programs for controlling the image restoration apparatus 100, and in particular, an image restoration program. For reference, the memory 120 may refer to a non-volatile storage device that maintains stored information even when power is not supplied, and a volatile storage device that requires power to maintain stored information.

The processor 130 controls the overall operation of the image reconstruction device 100. To this end, the processor 130 may include at least one processing unit (CPU, micro-processor, DSP, etc.), random access memory (RAM), read-only memory (ROM), CPU, graphic processing unit (GPU), and bus ( and a program stored in the memory 120 may be read into the RAM and executed through one or more processing units. Meanwhile, according to an embodiment, the term 'processor' may be interpreted to have the same meaning as terms such as 'control unit', 'controller', and 'computing device'.

The processor 130 executes a program stored in the memory 120, and processes the following operation by executing an image restoration program.

First, the processor 130 processes a feature extraction process for an input video file (hereinafter, referred to as an "input image file").

The processor 130 extracts image-based features from the video frames and constructs a feature vector representing the video frame.

The processor 130 extracts a predetermined kind of image features from damaged image data (eg, fragmented image frames included in an input image file). At this time, the processor 130 divides the image frame into a plurality of blocks based on vanishing points in the corresponding image frame, and extracts image characteristics in units of each divided block. The processor 130 generates a feature vector by vectorizing the extracted image features for each frame. As such, the extracted image characteristics may be expressed quantitatively by being represented by the characteristic vector. For reference, the image characteristic may include a color histogram and geometrical features (eg, edges and corners) to specify a corresponding frame.

The processor 130 divides each image frame included in the input image file into the smallest unit called a cell, and groups the divided cells into regions of a predetermined size (ie, blocks). In this case, the processor 130 may set an object of block region division to objects existing in the corresponding image frame. 'Object' means objects other than a background (eg, a vehicle, a lane, a sign, etc.) existing in an image frame. The processor 130 sets the unit of block region division to the number of cells included in the block, and each block includes a different number of cells.

In this case, the number of cells per block is determined by how far the object included in the block is from the vanishing point and the specific width and height obtained therefrom, which can be expressed by Equation 1 below.

<Equation 1>

In Equation 1, N _C denotes the number of cells included in the block, D _v denotes a distance from the vanishing point of the object included in the block, and P _w and P _h are each obtained through D _v . It means a specific width and height. The processor 130 may calculate P _w and P _h using an inverse perspective transformation technique used in the image reconstruction technique.

Using the distance D _v from the vanishing point of the frame calculated through Equation 1, the size of the block regions including the objects existing in the frame is calculated, and the frame is divided into the block regions. can do.

2 and 3, a method of processing the feature extraction process by the processor 130 will be described in more detail. For example, FIGS. 2 and 3 illustrate a result of processing of the feature extraction process by the image restoration apparatus 100 using the black box image in which the actual traffic accident scene is recorded.

In one embodiment of the present invention, the image frame means the smallest semantic unit of the image file. Therefore, since one image frame includes only information on one image, if there is no meta information of the image file, a correlation with another frame cannot be known. Accordingly, the processor 130 extracts an image-based feature to construct a feature vector representing a frame in order to understand the correlation between image frames without meta information.

In general, a video event data recorder (VEDR) such as a black box device records a continuous scene because the driving record of the vehicle is consistently recorded. Accordingly, similarities between image frames may be compared using global features such as color histograms. However, the global characteristic does not distinguish the background from the object, and there is a limitation in that it does not detect geometric changes caused by the movement of the object. In order to overcome such a problem, a complementary effect may be obtained by combining a global feature and a local feature. In addition, the video recorded by the black box device should be robust to the movement and rotation of the object, as vehicle driving is the main content. Accordingly, the processor 130 applies a FAST corner detection method to extract local features robust to the movement and rotation of the object.

2 is an example for describing a color characteristic and a geometric characteristic extraction method for each image frame block according to an embodiment of the present invention. 3 is an example for describing a vanishing point-based block generation method according to an embodiment of the present invention.

Since the black box device records more than 30 frames per second by default, the calculation for feature extraction increases exponentially with the recording time. In addition, one of the characteristics of the black box device is that the device is attached to a fixed position of the vehicle. Therefore, the vanishing point is at the same position in every frame, and since the geometric change of the object is proportional to the distance from the vanishing point while the object moves between frames, the geometric change of the object can be calculated with respect to the position.

Thus, as shown in FIG. 2, the processor 130 divides the image frame into cells, which are small pieces of a fixed size. These cells are used as the minimum unit of each vector in the calculation process. The processor 130 extracts the vanishing point from any frame in order to apply the characteristics of the vanishing point. The processor 130 generates blocks based on the vanishing points and then extracts the characteristics of each block.

Specifically, referring to FIG. 3, when an image frame as a raw image is input as in (a), the image frame is divided into a plurality of cells having the same size but connected to each other as in (b). . As shown in (c), a vanishing point is extracted from the corresponding image frame, and then blocks are generated based on the vanishing point as shown in (d).

If the VEDR is mounted in a fixed position, such as a black box device, one vanishing point applies to all frames. Thus, in order to reflect the absolute movement of the object irrespective of the distance D _V from the vanishing point, the processor 130 performs inverse perspective transformation based on the vanishing point. In this case, the processor 130 divides the calculated area into grids having a height P _h and a width P _w and then reflects the original area to the original area, and then includes cells included in each grid. Create a block by enclosing (N _C ).

As described above, the color characteristic is the most basic characteristic representing the frame image, but the disadvantage of not distinguishing the difference from the similar image is distinct, so that the processor 130 uses the geometric characteristics (ie, edge line information and edge information, etc.) together. To accurately recognize objects in the image area. As shown in FIG. 2, the feature extracted by the processor 130 may be a color histogram that is a global feature and edge information that is a local feature.

에 따라 기하 급수적으로 증가한다. 따라서, 히스토그램은 먼저 색상(color)를 빈(bin)의 수 N_b로 이산화하여 생성된다. 히스토그램의 크기(N_h)는 다음의 수학식 2와 같이 계산될 수 있다.In this case, the color histogram is configured for all pixels of each block, but the amount of calculation is the number of color channels.

Accordingly increases exponentially. Thus, the histogram is produced by first discretizing the color by the number N _b of bins. The size N _h of the histogram may be calculated as in Equation 2 below.

<Equation 2>

In addition, the processor 130 obtains corner information by applying a FAST corner detection technique to the entire image (ie, the entire frame). The edge information is used to determine whether the same object is detected in each image, and thus is composed of a frame unit different from a local property stored in blocks.

The processor 130 arranges the color histogram of each block and the edge information of the frame to form a multidimensional feature vector for each frame. The feature vector of each frame includes i blocks having j properties, which can be expressed by Equation 3 below.

<Equation 3>

Next, the processor 130 processes a visualization process based on the result of processing the feature extraction.

The multidimensional vector generated in the feature extraction step described above does not represent the relationship itself between frames. Therefore, in order to obtain a perceptually meaningful result, the processor 130 performs a dimension reduction process for visualizing the frame. In other words, in order to provide visually meaningful results while maintaining the relationship between the image frames, feature vectors are visualized using dimensional reduction techniques.

For reference, there are a number of dimensional reduction techniques and their respective characteristics are different. Therefore, a dimensional reduction technique considering characteristics of VEDR image data should be used. Since image data is a sum of discrete frames, it has a nonlinear structure but has a continuous context. That is, before and after frames contain similar image information. In addition, VEDR image data, such as a black box image, has a crowd problem that makes it difficult to find a difference between frames since the recorded image is almost the same during the period when the vehicle is stationary. Accordingly, the processor 130 processes dimension reduction using a t-distributed Stochastic Neighbor Embedding (t-SNE) technique in consideration of the characteristics of the VEDR image data.

In detail, the processor 130 maps the generated multi-dimensional feature vectors into a two-dimensional space. Each feature vector is composed of high dimensions because it contains all the image features of the frame. Therefore, similarity measurement between the feature vectors is possible, but there is a difficulty in determining the order of the frames. Accordingly, the processor 130 performs a dimension reduction process and a distribution measurement process on the feature vector through the t-SNE method and converts it into a 2D vector.

For reference, the t-SNE technique is a technique to compensate for the shortcomings of the SNE technique. First, the SNE technique will be described.

The process of mapping the high dimensional characteristic by the SNE technique to the low dimensional can be expressed as Equation 4 below.

 <Equation 4>

In Equation 4, p denotes the probability that the j th neighbor α _j is selected when the i th characteristic α _i existing in the high dimension space is given, and q is the i th entity y _i mapped to the low dimension. The probability that the j th neighbor y _j is selected. For reference, if the dimension reduction is normally performed, the probability of being selected as the neighbor in the high-dimensional space and the probability of being selected as the neighbor in the low-dimensional space is similar, so the purpose of the SNE is to minimize the distribution difference between p and q. The Kullback-Leibler divergence can be used as an indicator of how similar these two probability distributions are. According to the Coolback Librarian divergence calculation, if the two characteristic distributions of the high and low dimensions are completely different, the value is '1' and if the two characteristic distributions are the same, the value is '0'. Using this, the SNE proceeds with the dimension reduction to minimize the cost function of Equation 3 below.

<Equation 3>

The t-SNE technique, on the other hand, is a dimension reduction technique for nonlinear data, which preserves the local properties of the data manifold in the low dimensional representation and reflects the continuous context of the frame. The t-SNE technique is also robust to crowd problems by using student t-distribution. That is, the t-SNE technique uses a t-distribution rather than a Gaussian distribution used in the existing SNE to solve the crowd problem in the SNE technique described above.

FIG. 4 is an example of a result of visualizing a feature vector of an image frame by projecting on two-dimensional scattered points according to an embodiment of the present invention.

As shown in FIG. 4, the high-dimensional feature vector is reduced in two dimensions and visualized in a scatter plot graph. The two-dimensional graph of FIG. 4 is a graph in which high-dimensional characteristics are mapped to the two-dimensional space from the high-dimensional space. Here, the X-axis (ie, the horizontal axis) and the Y-axis (ie, the vertical axis) mean values that can be represented by the x and y value data shown in the probability distribution of p and q in the t-SNE technique. Since x is a high dimensional characteristic value and y is a low dimensional characteristic value to which a high dimensional element is mapped, the probability distribution calculated by the t-SNE technique becomes similar in the same frame and appears in a similar position on the 2D graph. That is, in FIG. 4, each point corresponds to each frame of the image file, and similar frames are disposed together in a closed state. Accordingly, the order of image frames can be intuitively recognized, and individual images can be classified into a curve-linear shape of a point cloud.

Next, the processor 130 processes a frame sequence reconstruction process based on a result of processing the visualization process.

The processor 130 reconstructs the original video file by reconstructing the frame order using the feature vectors that are dimensionally reduced and mapped to the lower dimension as described above. In this case, the processor 130 may configure the feature vectors reduced in two dimensions into a model graph by using a curve-fitting technique.

As described above, the feature vector of the low dimensional region maintains a gradually changing shape at higher dimensions due to the continuous context of the VEDR image. Accordingly, the visualized result of the two-dimensional graph has a curved-shaped distribution. This means that the feature vector of the frame is gradually changing, that is, the flow of the image.

Accordingly, the processor 130 may construct an appropriate sequence of images by finding a representative model suitable for the visualized curved distribution. Principal curve technique is one of nonlinear curve fitting techniques suitable for the distribution of feature vectors in VEDR images. The main curve technique is defined as self-consistent smooth curves that pass through the middle of the data cloud.

5 is an example of a result of visualizing a two-dimensional feature vector according to an embodiment of the present invention through a curve-fitting technique.

Referring to FIG. 5, the distribution of the two-dimensional feature vector may be fitted to a standard model of an image sequence by using a curve fitting technique. In this case, using the given curve fitting model, the sequence may be reconstructed by moving from the beginning to the end of the curve and calculating the nearest frame to reconstruct the sequence.

In this case, the processor 130 calculates orthogonal positions of the feature vectors and the model graph, and considers the point as the point where the frame of the vector is located in the order of the entire image. That is, the processor 130 may extract the order of the corresponding frame from the entire image based on the orthogonal positions of each two-dimensional vector and the model graph.

As shown in FIG. 4, the two-dimensional vectors show directionality according to the time order of the frames. By applying the curve fitting technique to the two-dimensional vectors based on the directionality, it is possible to generate a model graph for the driving image. That is, by generating the model graph shown in Figure 5, through this it is possible to restore the complete image by detecting the order of the frame.

1 again, the output module 140 outputs a result (ie, an image in which the frame order is restored) of the various processes of the processor 130 described above. In this case, the output module 140 may include a function of a communication module that transmits data (ie, restored image) by performing communication with an external device (ie, a black box device, etc.) to which the interworking external device is connected. In addition, the output module 140 may output the restored image through an output device (not shown) included in the image restoration apparatus 100 itself.

Meanwhile, the processor 130 of the image reconstruction device 100 according to another embodiment of the present invention detects a specific event in the image using the cuff fitting model generated through the process described with reference to FIGS. 2 to 5. Can be.

For example, special events such as car accidents or sharp turns may occur while driving. Such an event may cause a sudden change in the image data of a frame, which is reflected in the relationship between the characteristic vectors. In this case, using the curve fitting model, the vector direction and the rate of change of each frame can be calculated. As a result of this calculation, performing a second derivative on the curve can identify the point where the abrupt change occurred or the suspect event point. In this case, whether or not the change is rapid may be determined by calculating adaptive thresholding. That is, since the state of the recorded image varies depending on the characteristics of the road or the flow of the surrounding vehicles, a variable binarization value is used.

In this case, the processor 130 may separately output information about a specific event detected in the image or an image frame portion including the corresponding event through the output module 140.

Hereinafter, an image restoration method through the image restoration apparatus 100 will be described with reference to FIG. 6.

6 is a flowchart illustrating an image restoration method according to an embodiment of the present invention.

As shown in FIG. 6, first, a feature extraction process is processed on an input image file (S610).

That is, image features are extracted from fragmented image frames of the input image file, wherein each image feature includes a color histogram and a geometric feature to specify the frame. In addition, for more effective feature extraction, a frame may be divided into a plurality of blocks based on a vanishing point extracted from a corresponding frame, and then the feature may be extracted in units of blocks.

Since the feature extraction process is the same as the process described with reference to FIGS. 2 to 3 above, a detailed description thereof will be omitted.

Next, a visualization process is processed based on the result of the feature extraction process (S620).

First, the characteristic vector generated in step S610 is mapped to the two-dimensional space. At this time, since each feature vector includes all the image features of the frame, the feature vector is composed of three or more dimensions. Therefore, similarity between vectors can be measured, but it is not easy to determine the frame order. Perform dimension reduction. This transforms the high-dimensional feature vectors into two-dimensional vectors.

Since the visualization process of the feature vector is the same as the process described with reference to FIG. 4 above, a detailed description thereof will be omitted.

Next, the frame order of the input image file is reconstructed based on the result of the visualization of the frame characteristics (S630).

 First, the frame order is restored using the transformed two-dimensional vector in step S620. As an example, the curve fitting technique is used to construct two-dimensional vectors into a model graph. Subsequently, orthogonal positions of the vectors and the model graph are calculated, and the calculated positions are processed in the order of the corresponding frames in the entire image to restore the order of each frame.

Since the image restoration process using the visualized image frame is the same as the process described with reference to FIG. 5 above, a detailed description thereof will be omitted.

Then, the reconstructed order of each image frame is outputted (S640).

As described above, the image restoration method according to an embodiment of the present invention may restore the order of frames even when there is no encoding information or metadata of the corresponding image file.

The image restoration method according to the embodiment of the present invention described above may be implemented in the form of a recording medium including instructions executable by a computer, such as a program module executed by a computer. Such recording media include computer readable media, and computer readable media can be any available media that can be accessed by a computer and includes both volatile and nonvolatile media, removable and non-removable media. In addition, computer readable media includes computer storage media, which are volatile and nonvolatile implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. , Both removable and non-removable media.

The above description of the present invention is intended for illustration, and a person of ordinary skill in the art may understand that the present invention can be easily modified in other specific forms without changing the technical spirit or essential features of the present invention. will be. Therefore, it should be understood that the embodiments described above are exemplary in all respects and not restrictive. For example, each configuration characteristic described as a single type may be implemented in a distributed manner, and similarly, configuration characteristics described as distributed may be implemented in a combined form.

In addition, while the methods and systems of the present invention have been described in connection with specific embodiments, some or all of their configuration characteristics or operations may be implemented using computer systems having a general purpose hardware architecture.

The scope of the present invention is shown by the following claims rather than the detailed description, and all changes or modifications derived from the meaning and scope of the claims and their equivalents should be construed as being included in the scope of the present invention. .

100: image restoration device
110: input module
120: memory
130: processor
140: output module

Claims (13)

In the image restoration method through the image restoration apparatus,
Generating a characteristic vector by extracting a predetermined type of image characteristic for each frame of the input image file;
Visualizing the feature vector by mapping it to a low dimensional space through a dimension reduction process;
Generating a model graph optimized for the distribution of the feature vectors according to the visualization, and reconstructing the order of the frames in the image based on the model graph and the feature vector for each frame; And
Restoring an image file based on the order of the reconstructed frames;
Generating the feature vector,
Dividing a frame of the input image file into a plurality of cells each having the same size and being consecutive;
Extracting vanishing points of the frame;
Grouping the divided plurality of cells into a plurality of blocks each having an arbitrary size based on the vanishing point;
Extracting an image characteristic of a predetermined type for each block; And
Vectorizing the extracted image characteristic to generate a characteristic vector.
delete The method of claim 1,
Wherein the image characteristic comprises a color histogram and a geometrical feature. The method of claim 1,
Grouping into the plurality of blocks,
Splitting a block based on at least one object included in the frame, the image reconstruction method for calculating the height and width for each block through an inverse perspective transformation process based on the distance between the object and the vanishing point . The method of claim 1,
The visualizing step,
performing a dimension reduction process of mapping the characteristic vector to a two-dimensional space using a t-distributed Stochastic Neighbor Embedding (t-SNE) technique; And
And outputting a scatter plot graph of the two-dimensional vectors generated as a result of the dimension reduction process. The method of claim 5, wherein
Reconstructing the order of the frames,
Generating the two-dimensional vectors into a model graph using a curve-fitting technique;
Calculating orthogonal positions of the two-dimensional vectors and the model graph, and detecting the calculated positions in the order of the corresponding frames in the entire image; And
And reconstructing the order for each frame of the input image file based on the detected order. In the image restoration apparatus,
A memory in which an image restoration program is stored; And
A processor for executing a program stored in the memory;
The processor extracts a predetermined type of image characteristic for each frame of the input image file according to the execution of the image reconstruction program, generates a characteristic vector, maps the characteristic vector to a low dimensional space through dimensional reduction processing, and visualizes it. Generating a model graph optimized for the distribution of the feature vectors according to the visualization, reconstructing the order within the images of the frames based on the model graph and the feature vector for each frame, and based on the reconstructed frames Recover the file,
The processor,
The frame of the input image file is divided into a plurality of consecutive cells having the same size, respectively,
Grouping the plurality of divided cells into a plurality of blocks each having an arbitrary size, based on a vanishing point extracted from the frame,
Extracting a predetermined kind of image characteristic for each block;
And reconstructing the extracted image characteristic to generate a characteristic vector. delete The method of claim 7, wherein
Wherein the image characteristic comprises a color histogram and a geometrical feature. The method of claim 7, wherein
The processor,
Partition the block based on at least one object included in the frame,
And calculating the height and width for each block through an inverse perspective transformation process based on the distance between the object and the vanishing point, and grouping the frame into the plurality of blocks. The method of claim 7, wherein
The processor,
a dimension reduction process that maps the characteristic vector to two-dimensional space using a t-distributed Stochastic Neighbor Embedding (t-SNE) technique,
And a scatter plot graph of the two-dimensional vectors generated as a result of the dimension reduction process to visualize the characteristic vector of the frame. The method of claim 11,
The processor,
The two-dimensional vectors are generated as a model graph using a curve-fitting technique,
Calculating orthogonal positions of the two-dimensional vectors and the model graph, detecting the calculated positions as the order of the corresponding frames in the entire image,
And reconstruct the order for each frame of the input image file based on the detected order. A computer-readable recording medium having recorded thereon a program for implementing the method of claim 1.

Images