KR20180071947A - Apparatus and method for processing image - Google Patents

Abstract

The present invention relates to an image processing apparatus and method. The image processing apparatus according to the present embodiment includes a feature extraction part for extracting time information by applying a convolution filter to an input frame, a patch dividing part for dividing a feature map including the time information into pixel units to acquire a feature patch, a temporal information processing part for extracting temporal information on the feature patch, a probability classifying part for extracting one probability result for each pixel from the temporal information, and a background separation part for separating the foreground and background of the input frame based on the extracted probability result.

Description

[0001] APPARATUS AND METHOD FOR PROCESSING IMAGE [0002]

Embodiments disclosed herein relate to an image processing apparatus and method. More particularly, the present invention relates to an image processing apparatus and method for separating foreground and background of a moving image by using an artificial neural network.

In the conventional image processing method, the input image is compared with the previous image and the current image in order to separate the foreground and background from each other. Korean Patent No. 10-2013-0063963, which is related to the prior art, compares the current image with the previous image, accumulates the change value between the images, generates the accumulated image change value, and uses the accumulated image change value Describes an image processing method for estimating a change value of an image by moving an object.

Conventional image processing methods including this method set the front part of the moving picture as a learning section of the background model so that all regions are divided into a background model without separating the foreground and the foreground and background are separated from the input image after the learning section, And so on. At this time, various techniques are used to reduce noise and improve performance in background learning, separation, and update. The user directly designes for noise reduction and uses different parameters for each image characteristic. Because of this, all parts for image processing must be designed by the user, complicated memory update operation is performed, and the speed can be limited because it is mainly focused on CPU operation.

In this way, existing image processing methods are not easy to implement because they are directly designed by a user using a non-learning-based image processing method, and there is a limitation in performance due to inefficiency of use of hardware resources.

Therefore, in recent years, apparatuses and methods for solving such problems have been demanded.

On the other hand, the background art described above is technical information acquired by the inventor for the derivation of the present invention or obtained in the derivation process of the present invention, and can not necessarily be a known technology disclosed to the general public before the application of the present invention .

SUMMARY OF THE INVENTION [0008] Embodiments disclosed in the present application are directed to an image processing apparatus and method for separating a foreground and a background of a moving image by using an artificial neural network.

The embodiments disclosed herein are directed to an image processing apparatus and method for extracting temporal information of a moving image by using an artificial neural network and separating foreground and background using the extracted temporal information.

The embodiments disclosed in the present specification aim to provide an image processing apparatus and method capable of automatically learning, separating, or updating a background using an artificial neural network without user intervention.

It is an object of the present invention to provide an image processing apparatus and method capable of detecting motion information by learning an artificial neural network so as to utilize not only time information but also temporal information.

According to an embodiment of the present invention, there is provided an image processing apparatus including a feature extraction unit for extracting time information by applying a convolution filter to an input frame, , A temporal information processing unit for extracting temporal information with respect to the feature patch, a probability classification unit for extracting one probability result for each pixel from the temporal information, and a probability distribution unit And a background separator for separating the foreground and the background of the input frame.

According to another embodiment, an image processing method performed by an image processing apparatus includes extracting time information by applying a convolution filter to an input frame, dividing a feature map including the time information into pixels, Extracting temporal information for the feature patch, extracting one probability result for each pixel from the temporal information, and separating the foreground and background of the input frame based on the extracted probability result .

According to any one of the above-mentioned means for solving the above-mentioned problems, an image processing apparatus and method for separating the foreground and background of a moving image by using an artificial neural network can be presented.

In addition, according to any one of the above-mentioned problems, an image processing apparatus and method for extracting temporal information of a moving image by using an artificial neural network and separating foreground and background using the extracted temporal information can be presented.

Further, according to any one of the above-mentioned means for solving the above-mentioned problems, it is possible to provide an image processing apparatus and method capable of automatically learning, separating, or updating the background using the artificial neural network without user intervention.

In addition, according to any one of the above-mentioned problem solving means, an image processing apparatus and method capable of detecting motion information by learning an artificial neural network so as to utilize not only time information but also temporal information can be presented.

The effects obtainable in the disclosed embodiments are not limited to the effects mentioned above, and other effects not mentioned can be clearly understood from the following description, by those skilled in the art, .

1 is a block diagram illustrating an image processing apparatus according to an exemplary embodiment of the present invention.
FIG. 2 is a diagram for explaining a process of extracting time information from an input frame according to an embodiment.
FIG. 3 is a diagram for explaining a process of extracting a feature patch from a feature map according to an embodiment.
4 is a diagram for explaining a process of processing temporal information according to one embodiment.
5 is a diagram for explaining the internal structure of the circular neural network shown in FIG.
6 is a diagram illustrating a process of outputting a probability result on a pixel-by-pixel basis according to an exemplary embodiment.
7 is a flowchart illustrating an image processing method according to an embodiment.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings, which will be readily apparent to those skilled in the art. The present invention may, however, be embodied in many different forms and should not be construed as limited to the embodiments set forth herein. In order to clearly illustrate the present invention, parts not related to the description are omitted, and similar parts are denoted by like reference characters throughout the specification.

Throughout the specification, when a part is referred to as being "connected" to another part, it includes not only "directly connected" but also "electrically connected" with another part in between . Also, when an element is referred to as "comprising ", it means that it can include other elements as well, without departing from the other elements unless specifically stated otherwise.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, the present invention will be described in detail with reference to the accompanying drawings.

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

1, the image processing apparatus 100 includes a feature extracting unit 110, a patch dividing unit 120, a temporal information processing unit 130, a probability classifying unit 140, and a background separating unit 150, .

The feature extraction unit 110 may extract the time information that can be composed of feature maps from the input frame. Here, the input frame may be one of a plurality of frames constituting a moving picture.

The feature map extracted by the feature extraction unit 110 includes time information for at least a part of an input frame. The feature extraction unit 110 may use a convolution filter to acquire the feature map. Here, the convolution filter may be implemented using a Convolution Neural Network (CNN). Accordingly, the feature extraction unit 110 can learn the convolution filter necessary for background separation using the CNN from the input frame, i.e., the RGB image, and extract the feature map using the learned convolution filter .

The patch dividing unit 120 may divide the feature map into pixel units to obtain a feature patch. When the feature map is composed of a plurality of layers, the patch dividing unit 120 can divide the feature map into feature patches including all of the time information for each of the plurality of layers. Since the patch dividing unit 120 divides the feature map in units of pixels, it can acquire a plurality of feature patches from one feature map.

The temporal information processing unit 130 can extract temporal information for each feature patch. Here, the temporal information processing unit 130 may use a Recurrent Neural Network (RNN) for extracting temporal information. At this time, the temporal information processing unit 130 extracts the temporal information included in the moving picture by calculating it together with the current feature patch using the information extracted from the previous feature patch. In addition, the temporal information processing unit 130 can update the internal state of the Cyclic Neural Network, that is, the cell state.

The probability classifying unit 140 may extract a probability result for each pixel from the temporal information. For this, the probability classifier 140 may use a Fully Connected Neural Network (FCNN).

The background separating unit 150 may detect a motion for a specific object in the input frame based on the extracted probability result. Accordingly, the background separator 150 can separate the foreground and the background from each other in the input frame.

As described above, the proposed image processing apparatus 100 can separate images using an artificial neural network. As described above, the image processing apparatus 100 may use a structure in which artificial neural networks of CNN, RNN, and FCNN are sequentially combined to detect motion of a specific object in a moving image. In particular, the image processing apparatus 100 enables the learning of the artificial neural network that utilizes the temporal information without using the visual information of the moving picture by using the RNN structure.

Further, the image processing apparatus 100 can automatically learn, separate, or update the background without using the user's direct design using the artificial neural network, and the performance change due to parameter selection, etc. is small.

FIG. 2 is a diagram for explaining a process of extracting time information from an input frame according to an embodiment.

As shown in FIG. 2, the feature extraction unit 110 may extract the time information from an input frame using a convolution filter, and may use CNN for this purpose.

The feature extraction unit 110 can use the convolution filter 220 for a partial area 211 of the input frame 210 and can extract the time information (or CNN feature) that can be composed of the feature map 230 Can be extracted. 2, the feature map 230 is shown to include four layers 231, 232, 233, and 234, but the feature map 230 is not limited to having one or more layers using a convolution filter And can be set to various numbers.

The feature extraction unit 110 may learn a convolution filter necessary for background separation from the input image 210. [ Here, the feature extraction unit 110 may learn and generate a plurality of feature maps 230 by performing a role of a filter bank with a convolution filter. Accordingly, the feature extraction unit 110 can automatically extract and extract information necessary for background separation.

Meanwhile, the CNN described above can support various functions for searching, classifying, and understanding images for image processing through learning of images as one of artificial neural networks. Because CNN has a different structure from other artificial neural networks that learn by connecting all areas of input, one parameter can be used in various areas of the input frame. Therefore, the CNN can process the image with a small number of parameters, and even if the image is spatially different, the characteristics are not changed, so that the image analyzing apparatus 100 can be used for extracting the visual information.

FIG. 3 is a diagram for explaining a process of extracting a feature patch from a feature map according to an embodiment.

As shown in FIG. 3, the patch dividing unit 120 can perform patch division for each of the plurality of layers 231, 232, 233, and 234 through patch division of the feature map 230. FIG. In this way, the patch dividing unit 120 can obtain the feature patch 240 in units of pixels from the feature map 230. [ Here, the feature patch 240 may include time information on a pixel-by-pixel basis of each of the plurality of layers 231, 232, 233, and 234. In addition, the number of layers of the feature map 230 may be divided into L (L = 4), and L is a natural number of 1 or more.

The patch dividing unit 120 may generate a plurality of feature patches in units of pixels by dividing the feature map 230. [ In addition, the patch dividing unit 120 may divide the image into a plurality of strides in consideration of image processing speed.

On the other hand, as the number L of layers of the feature map 230 divided by the patch dividing unit 120 increases, more time information for separating foreground and background can be included. As a result, if the number of layers in the feature map 230 increases, the performance according to the separation of the foreground and the background can be improved.

4 is a diagram for explaining a process of processing temporal information according to one embodiment.

As shown in FIG. 4, the temporal information processing unit 130 may extract the temporal feature 260 from the feature patch 240. Here, the temporal information processing unit 130 may extract the temporal feature 260, which is a prediction value for separating the foreground and the background, from the temporal information flow as time elapses. Here, the temporal information processing unit 130 may be implemented using the RNN 250. [

Also, the temporal information processing unit 130 may output the temporal information 260 on a pixel-by-pixel basis, and may receive the cell state of the previous input as the next input. Accordingly, the temporal information processing unit 130 can extract the temporal information 260, which is a predictive value for the separation of the foreground and the background, using the feature vector extracted from the feature patch 240 in units of pixels.

5 is a diagram for explaining the internal structure of the circular neural network shown in FIG.

As shown in FIG. 5, the RNN 250 is a neural network configured to feed back information on previous inputs by using an internal loop for a series of sequential inputs. Accordingly, the RNN 250 can perform a different operation for each input by using the internal state information for the sequential input, and learn the matrix for updating the cell state information through learning.

The RNN 250 may be implemented as a short-term memory (LSTM). The LSTM RNN 250 includes a cell state operation line 251, a Forget Gate Layer 252, an input gate layer 253, An Update Cell State) 254, and an Output Gate Layer 255. The RNN 250-1 and the RNN 250-3 may have a similar structure to the RNN 250-2, and may perform operations before and after the RNN 250-2. Respectively.

Cell state line 251, it contains a multiply operation and add operations, receives the C _{t- 1} was previously output to the internal state (that is, the cell state) may output a C _t is the current cell state. The cell state 251 determines whether or not the time information (i.e., the time information included in the feature patch) is reflected using the getgate gate layer 252, the input gate layer 253, and the output gate layer 255 .

) 연산을 이용하여, h_t-1과 x_t를 입력으로 수신한다. 여기서, x_t는 현재 프레임의 특징 패치(240)이다. 따라서, x_{t -1}은 이전 프레임의 특징 패치이고, x_{t +1}은 다음 프레임의 특징 패치이다. 또한, h_{t- 1}는 히든 스테이트로서, 이전 RNN(250-1)의 출력값을 나타낸다. 포겟 게이트 레이어(252)는 시그모이드 연산을 사용하여 0에서 1사이의 값을 출력(0<

<1)할 수 있으며, 시그모이드 연산의 출력값은 1에 가까울수록 정보의 반영을 많이 하고, 0에 가까울수록 정보의 반영을 적게하는 것을 의미한다.The getgate gate layer 252 includes a sigmoid

) Operation to receive h _t-1 and x _t as inputs. Here, x _t is the feature patch 240 of the current frame. Thus, x _{t -1} is the feature patch of the previous frame, and x _{t +1} is the feature patch of the next frame. Also, _{ht- 1} is a hidden state and represents the output value of the previous RNN 250-1. The getgate gate layer 252 outputs a value between 0 and 1 using a sigmoid operation (0 <

<1), and the output value of the sigmoid operation is closer to 1 to reflect more information, and closer to 0 means less information to be reflected.

The operation of the getgate gate layer 252 can be expressed by Equation (1) below.

Where f _t represents the output of the getgate gate layer 252. W _f is the weight of the foregget gate, b _f is the bias value of the getgate gate, and W _f and b _f are the values set by the learning of the RNN.

) 연산과 쌍곡 탄젠트(tanh) 연산을 포함할 수 있다. 여기서, 시그모이드 연산은 어떤 값을 업데이트할지를 결정하며, 쌍곡 탄젠트 연산은 셀 상태에 더해질 후보값들의 벡터를 만든다. 이때, 입력 게이트 레이어(253)에서 수행되는 시그모이드 연산과 쌍곡 탄젠트 연산 각각은 하기의 수학식 2와 같이 나타낼 수 있다.Next, the input gate layer 253 determines whether the new information is stored in the cell state. The input gate layer 253 includes a sigmoid

) Operation and a hyperbolic tangent (tanh) operation. Here, the sigmoid operation determines which value to update and the hyperbolic tangent operation produces a vector of candidate values to be added to the cell state. At this time, the sigmoid operation and the hyperbolic tangent operation performed in the input gate layer 253 can be expressed by the following Equation (2).

Where i _t is the output of the sigmoid operation in the input gate layer 253 and c _t is the output of the hyperbolic tangent operation in the input gate layer 253. Here too, W _i is the weight of the sigmoid operation and W _c is the weight of the hyperbolic tangent operation. Also, b _i is the bias value of the sigmoid operation and b _c is the bias value of the hyperbolic tangent operation. W _i , W _c , b _i , and b _c are values set by the learning of the RNN.

The update cell state 254 updates the values output from the get gate layer 252 and the input gate layer 253 to cell states C _t-1 and C _t . The update of the cell state can be expressed by the following equation (3).

Where C _t represents the cell state, and C _{t- 1} represents the previous cell state. f _t is the output of the getgate layer 252, i _t is the output of the sigmoid operation in the input gate layer 253 and c _t is the output of the hyperbolic tangent operation in the input gate layer 253.

Next, the output gate layer 255 determines the output value h _t . h _t is the filtered value of the cell state (C _t ). The output gate layer 255 may include a sigmoid operation and a hyperbolic tangent operation. Here, the sigmoid operation can calculate cell state determination information (o _t ) for determining a part of the cell state. The hyperbolic tangent operation multiplies the updated cell state (C _t ) by the hyperbolic tangent operation and the cell state determination information (o _t ) output from -1 to 1. The sigmoid operation and the hyperbolic tangent operation performed in the output gate layer 255 can be expressed by Equation (4) below.

O _t is the output of the sigmoid operation in the output gate layer 255 and h _t is the output of the hyperbolic tangent operation in the output gate layer 255. Here too, W _o is the weight of the sigmoid operation and b _o is the bias value of the sigmoid operation. W _o and b _o are values set by the learning of the RNN.

6 is a diagram illustrating a process of outputting a probability result on a pixel-by-pixel basis according to an exemplary embodiment.

As shown in FIG. 6, the probability classifier 140 can classify the probability 280 for foreground and background with respect to the temporal information 260, and performs a computation on a pixel-by-pixel basis. Can be obtained. For this, the probability classifier 140 may classify probabilities using the FCNN 270.

The probability classifier 140 applies the same parameters to the FCNN 270 for each pixel in the RNN 250 so that the use of parameters for separating foreground and background can be minimized. For example, when the FCNN 270 is applied to all the pixels, the probability classification unit 140 needs parameters proportional to the resolution, but the parameters used for each pixel in the RNN 250 are the same in the FCNN 270 The same parameters can be used regardless of the resolution.

As described above, since the probability classifier 140 uses the RNN 250 that outputs successive results for successive inputs, it interprets the temporal information represented as a continuous array of frames in addition to the visual information of the moving image can do.

7 is a flowchart illustrating an image processing method according to an embodiment.

As shown in FIG. 7, the image processing apparatus 100 can extract the time information that can be composed of the feature map from the input frame (S310). The image processing apparatus 100 can extract the time information in the form of a feature map and extract the feature map using a convolution filter in the input frame. Here, the feature map may be defined as a CNN feature by using a convolution filter implemented using CNN.

The image processing apparatus 100 can obtain a feature patch by dividing the feature map on a pixel-by-pixel basis (S320). At this time, when the feature map is composed of a plurality of layers, the image processing apparatus 100 can be divided into feature patches including all of the time information for each of the plurality of layers.

The image processing apparatus 100 may extract the temporal information using the RNN for each feature patch (S330). The image processing apparatus 100 is, for external memory (e.g., can be represented by h _t-1 is the output value of the RNN in Fig. 6, h _t, h _{t + 1,} and so on) and an internal memory (for example, in Fig. 6 (Which may be represented by cell states C _t-1 , C _t , C _{t + 1} , and so forth) to update the memory corresponding to the temporal flow, Can be extracted. In this way, the image processing apparatus 100 is, for feature vectors (such as time information, the weights in Figure 6 (W _f, W _i, W _c, W _o) and a bias (b _f, b _i, b _c, b _o ) can be extracted.

The image processing apparatus 100 does not depend only on the time information using the RNN, and can learn the neural network that can utilize the temporal information. Accordingly, the image processing apparatus 100 can detect motion information on a moving object. Also, the image processing apparatus 100 can store the past values obtained from the image frame using the RNN, and predict the next value.

The image processing apparatus 100 may extract a probability result for each pixel from the temporal information (S340). The image processing apparatus 100 can extract the probability result from the temporal information using the FCNN.

The image processing apparatus 100 can detect the motion of the object in the input image using the extracted temporal information (S350).

The image processing apparatus 100 may separate the foreground and the background through motion detection in the image and terminate the process (S360).

Meanwhile, the operation of the image processing apparatus 100 shown in FIG. 7 can separate images by sequentially using CNN, RNN, and FCNN. When the process shown in FIG. 7 is used, It can also be used for learning motion, and can also be used for actual image analysis operations to separate the actual foreground and background. Here, the image processing apparatus 100 can learn parameters in CNN, RNN, and FCNN during a learning operation or an actual image analysis, and can update parameters using the learned parameters.

The image processing apparatus 100 according to the present embodiment can perform a background difference (e.g., motion detection), which is one of automatic analysis techniques for moving images. Accordingly, the image processing apparatus 100 can detect a moving object in the moving image, reduce the area for analysis to a region where the moving object exists, and improve the image analysis speed. Also, the image processing apparatus 100 can acquire information on the presence / absence, direction, and type of motion in the moving image. Therefore, the image processing apparatus 100 can be applied to a surveillance field using images such as security and expense, and can be applied to various other fields where functions such as moving image synthesis, object recognition, and image retrieval can be utilized.

The term &quot; part &quot; used in the present embodiment means a hardware component such as software or a field programmable gate array (FPGA) or an ASIC, and 'part' performs certain roles. However, 'part' is not meant to be limited to software or hardware. &Quot; to &quot; may be configured to reside on an addressable storage medium and may be configured to play one or more processors. Thus, by way of example, 'parts' may refer to components such as software components, object-oriented software components, class components and task components, and processes, functions, , Subroutines, segments of program patent code, drivers, firmware, microcode, circuitry, data, databases, data structures, tables, arrays, and variables.

The functions provided within the components and components may be combined with a smaller number of components and components or separated from additional components and components.

In addition, the components and components may be implemented to play back one or more CPUs in a device or a secure multimedia card.

Also, the image processing method according to an embodiment of the present invention may be implemented as a computer program (or a computer program product) including instructions executable by a computer. A computer program includes programmable machine instructions that are processed by a processor and may be implemented in a high-level programming language, an object-oriented programming language, an assembly language, a machine language, or the like , In particular, R language, Python language, Ruby language, Scheme language, and the like. The computer program may also be recorded on a computer readable recording medium of a type (e.g., memory, hard disk, magnetic / optical medium or solid-state drive).

Therefore, the image processing method according to the embodiment of the present invention can be realized by the computer program as described above being executed by the computing device. The computing device may include a processor, a memory, a storage device, a high-speed interface connected to the memory and a high-speed expansion port, and a low-speed interface connected to the low-speed bus and the storage device. Each of these components is connected to each other using a variety of buses and can be mounted on a common motherboard or mounted in any other suitable manner.

Where the processor may process instructions within the computing device, such as to display graphical information to provide a graphical user interface (GUI) on an external input, output device, such as a display connected to a high speed interface And commands stored in memory or storage devices. As another example, multiple processors and / or multiple busses may be used with multiple memory and memory types as appropriate. The processor may also be implemented as a chipset comprised of chips comprising multiple independent analog and / or digital processors.

The memory also stores information within the computing device. In one example, the memory may comprise volatile memory units or a collection thereof. In another example, the memory may be comprised of non-volatile memory units or a collection thereof. The memory may also be another type of computer readable medium such as, for example, a magnetic or optical disk.

And the storage device can provide a large amount of storage space to the computing device. The storage device may be a computer readable medium or a configuration including such a medium and may include, for example, devices in a SAN (Storage Area Network) or other configurations, and may be a floppy disk device, a hard disk device, Or a tape device, flash memory, or other similar semiconductor memory device or device array.

It will be understood by those skilled in the art that the foregoing description of the present invention is for illustrative purposes only and that those of ordinary skill in the art can readily understand that various changes and modifications may be made without departing from the spirit or essential characteristics of the present invention. will be. It is therefore to be understood that the above-described embodiments are illustrative in all aspects and not restrictive. For example, each component described as a single entity may be distributed and implemented, and components described as being distributed may also be implemented in a combined form.

The scope of the present invention is defined by the appended claims rather than the detailed description and all changes or modifications derived from the meaning and scope of the claims and their equivalents are to be construed as being included within the scope of the present invention do.

100: Image processing apparatus 110: Feature extraction unit
120: patch division unit 130: temporal information processing unit
140: probability classification section

Claims (14)

A feature extraction unit for extracting time information by applying a convolution filter to an input frame;
A patch division unit for dividing a feature map including the time information by pixels to acquire a feature patch;
A temporal information processing unit for extracting temporal information on the feature patches;
A probability classifier for extracting one probability result for each pixel from the temporal information; And
And a background separator for separating the foreground and background of the input frame based on the extracted probability result. The method according to claim 1,
Wherein the patch dividing unit comprises:
Wherein when the feature map includes a plurality of layers, the feature patch includes pixel information of each of the plurality of layers. The method according to claim 1,
Wherein the temporal information processing unit comprises:
And extracting the temporal information using a circular neural network (RNN). The method of claim 3,
Wherein the temporal information processing unit comprises:
Wherein the temporal information is information including information on the flow of time information as time elapses, and extracts the temporal information as a predictive value for separating the foreground and the background. The method according to claim 1,
The convolution filter comprises:
An image processing apparatus that is implemented using a convolutional neural network (CNN). The method according to claim 1,
The probability classifying unit may classify,
And extracts the probability result using the total connected neural network (FCNN) from the temporal information. The method according to claim 1,
Wherein the background separator comprises:
Detecting an object motion in the input frame, and separating a foreground and a background through the motion detection. An image processing method performed by an image processing apparatus,
Extracting time information by applying a convolution filter to an input frame;
Obtaining a feature patch by dividing a feature map including the time information on a pixel basis;
Extracting temporal information for the feature patch;
Extracting one probability result for each pixel from the temporal information; And
And separating the foreground and background of the input frame based on the extracted probability result. 9. The method of claim 8,
Wherein acquiring the feature patch comprises:
And when the feature map includes a plurality of layers, dividing the feature patch so as to include all pixel information of each of the plurality of layers. 9. The method of claim 8,
The step of extracting the temporal information comprises:
And extracting the temporal information using a circular neural network (RNN). 11. The method of claim 10,
Wherein the temporal information comprises:
Wherein the information includes information on the flow of time information according to the flow of time, and is a predictive value for separating foreground and background. 9. The method of claim 8,
Wherein the extracting of the feature comprises:
And extracting features using the convolution filter implemented using a convolutional neural network (CNN). 9. The method of claim 8,
The step of extracting the probability result includes:
And extracting the probability result from the temporal information using a total connected neural network (FCNN). 9. The method of claim 8,
The step of separating the foreground and background comprises:
Detecting movement of an object within the input frame,
And separating foreground and background through the motion detection.

Images