KR20130110417A - Method for analyzing video stream data using multi-channel analysis - Google Patents

Abstract

PURPOSE: A video stream analysis method of using a multi channel analysis is provided to analyze a dependency relation of a division and a semantic boundary of a video stream by learning the video stream by classifying the video stream into an image channel and a sound channel and by referencing a likelihood change based on the learned data. CONSTITUTION: A video stream analysis method comprises the following steps: a step aims to classify a video stream data into an image channel and a sound channel, and to comprise learning models about the classified image channel and the classified sound channel respectively (S1010, S1020); a step (S1030) aims to estimate a likelihood of a current frame set about a frame set which follows by using the learning models of the comprised image channel and the comprised sound channel; and a step (S1040) aims to record trends of the estimated likelihood, and to decide a story transition point of a video stream by using a peak point exceeding a predetermined threshold among the trends of a recent recorded likelihood. [Reference numerals] (AA) Start; (BB) End; (S1010) Classify a video stream data into an image channel and a sound channel; (S1020) Comprise learning models about the classified image channel and the classified sound channel respectively; (S1030) Estimate a likelihood of a current frame set about a frame set which follows by using each learning model; (S1040) Decide a story transition point by using a peak point exceeding a predetermined threshold

Description

Video stream analysis method using multi channel analysis {METHOD FOR ANALYZING VIDEO STREAM DATA USING MULTI-CHANNEL ANALYSIS}

The present invention relates to an analysis technique for a video stream. More specifically, the video stream is divided into an image channel and a sound channel, respectively, and the semantic boundary of the video stream is referred to based on the likelihood change based on the learned data. The present invention relates to a video stream analysis method using multi-channel analysis that can classify and analyze.

With the development of imaging technology, various analyzes of video data are being performed. In particular, various attempts have recently been made to perform semantic classification on video streams.

In the past, however, it merely detected a change in the expression (composition) of the screen to distinguish or compare images. For example, the analysis of screen switching and the like using only changes in RBG (Red-Blue-Green) values of an image has been performed, and the actual semantic classification cannot be successfully achieved.

The present invention classifies a video stream into an image channel and a sound channel, respectively, and uses multi-channel analysis to analyze a dependency relationship of semantic boundaries of a video stream by referring to a likelihood change based on the learned data. It is intended to provide a video stream analysis method.

In addition, the present invention provides a video stream analysis method using multi-channel analysis that can learn the image channel flexibly and more accurately by performing hierarchical learning when learning the image channel is not limited to a specific feature to learn I would like to.

Among the embodiments, the video stream analysis method is performed in a video strip analysis apparatus capable of receiving a video stream and analyzing the story transformation of the received video stream. The video stream analysis method includes (a) dividing video stream data into an image channel and a sound channel, and constructing learning models for each of the divided image channel and sound channel, and (b) configuring the image channel and sound channel. Estimating the likelihood of the current frame set with respect to the trailing frame set using learning models; and (c) recording the estimated likelihood trend and a peak point exceeding a predetermined threshold value among the recorded likelihood trends. Determining a story transformation point of the video stream by using.

In one embodiment, the step (a) includes separating the video stream data into the image channel and the sound channel, and constructing a learning model of the image channel using visual words for the separated image channel. can do.

In an embodiment, the constructing the learning model of the image channel may include extracting the visual word by applying a Scale Invariant Feature Transform (SIFT) to the existing image channel.

In an embodiment, the step (a) may include extracting a feature using a Mel Frequency Cepstral Coefficients (MFCC) algorithm for the previously separated sound channel.

In one embodiment, the learning model of the sound channel may perform modeling based on the continuity of the dialogue reflected in the existing sound channel.

In one embodiment, the step (c) is a second peak point of the likelihood estimated from the learning model of the sound channel within a predetermined time range around a first peak point of the likelihood estimated from the learning model of the previous image channel. If present, the method may include determining the first peak point as the story transformation point.

In one embodiment, the step (c) is a step of determining the conversion point using a sequential hierarchical Dirichlet process (sHDP) of dividing the image channel and the sound channel into a lower stream, respectively; It may include.

Among the embodiments, the video strip analyzing apparatus may receive a video stream and analyze a story transformation of the received video stream. The video strip analyzing apparatus includes a separation module, an image learning module, a sound learning module and a control module. The separation module separates the video stream into an image stream and a sound stream. The image learning module extracts a visual word for the separated image stream and uses the same to estimate a likelihood of the current frame set with respect to the following frame set. The sound learning module performs modeling on the separated sound stream based on continuity of dialogue, and estimates the likelihood of the current frame set with respect to the following frame set using the generated model. The control module identifies a peak point exceeding a preset threshold based on the likelihood of the likelihood estimated by the image learning module and the sound learning module, and associates the peak points of the image stream and the sound stream with respect to the video stream. Determine story change points.

In one embodiment, the control module may determine the conversion point using a serial Hierarchical Dirichlet Process (sHDP) that divides and processes the image channel and the sound channel into sub streams, respectively.

In example embodiments, the control module may be configured to, when the second peak point of the likelihood estimated by the sound learning module is present within a predetermined time range based on the first peak point of the likelihood estimated by the image learning module. The peak point may be determined as the story change point.

Among the embodiments, the recording medium records a program for executing the video stream analysis method. The program may be driven in a video strip analysis apparatus that receives a video stream and analyzes the story transformation of the received video stream. (A) The video stream data is divided into an image channel and a sound channel, and the classification is performed. (B) estimating the likelihood of the current frame set with respect to the following frame set using the configured image models and the learning channel of the sound channel, and (c) Recording the estimated likelihood trend and determining a story conversion point of the video stream by using a peak point exceeding a preset threshold value among the recorded likelihood trends.

According to the present invention, the video stream is divided into an image channel and a sound channel, respectively, and each of them is trained, and the dependency relationship between the semantic boundaries of the video stream can be analyzed with reference to the likelihood change based on the learned data. .

In addition, according to the present invention, it is possible to learn the image channel flexibly and more accurately by performing hierarchical learning without limiting to specific features when learning the image channel.

1 is a reference diagram schematically illustrating a data processing schema according to the present invention.
2 is a reference diagram illustrating an algorithm for sHDP modeling according to the present invention.
3 is a reference diagram for explaining an sHDP model.
4 to 6 are graphs showing trends of recall, precision and correct answers of episodes 1 to 3, respectively.
7 is a graph showing the F1 measure in episodes 1-3.
8 is a graph showing the effect of the channel merging method according to the present invention.
9 is a block diagram illustrating an embodiment of a video strip analysis apparatus according to the present invention.
10 is a flowchart illustrating an embodiment of a video strip analysis method according to the present invention.

The description of the present invention is merely an example for structural or functional explanation, and the scope of the present invention should not be construed as being limited by the embodiments described in the text. That is, the embodiments are to be construed as being variously embodied and having various forms, so that the scope of the present invention should be understood to include equivalents capable of realizing technical ideas.

Meanwhile, the meaning of the terms described in the present invention should be understood as follows.

The terms "first "," second ", and the like are intended to distinguish one element from another, and the scope of the right should not be limited by these terms. For example, the first component may be referred to as a second component, and similarly, the second component may also be referred to as a first component.

It is to be understood that when an element is referred to as being "connected" to another element, it may be directly connected to the other element, but there may be other elements in between. On the other hand, when an element is referred to as being "directly connected" to another element, it should be understood that there are no other elements in between. On the other hand, other expressions describing the relationship between the components, such as "between" and "immediately between" or "neighboring to" and "directly neighboring to", should be interpreted as well.

It should be understood that the singular " include "or" have "are to be construed as including a stated feature, number, step, operation, component, It is to be understood that the combination is intended to specify that it does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, or combinations thereof.

In each step, the identification code (e.g., a, b, c, etc.) is used for convenience of explanation, the identification code does not describe the order of each step, Unless otherwise stated, it may occur differently from the stated order. That is, each step may occur in the same order as described, may be performed substantially concurrently, or may be performed in reverse order.

The present invention can be embodied as computer-readable code on a computer-readable recording medium, and the computer-readable recording medium includes all kinds of recording devices for storing data that can be read by a computer system . Examples of the computer-readable recording medium include a ROM, a RAM, a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, and the like, and also implemented in the form of a carrier wave (for example, transmission over the Internet) . In addition, the computer-readable recording medium may be distributed over network-connected computer systems so that computer readable codes can be stored and executed in a distributed manner.

All terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs, unless otherwise defined. Commonly used predefined terms should be interpreted to be consistent with the meanings in the context of the related art and can not be interpreted as having ideal or overly formal meaning unless explicitly defined in the present invention.

1. Sequential HDP

Episodes of TV dramas consist of several story scenes. Each story section consists of a series of frames (image channels) and at the same time a set of dialogues (sound channels). Each frame consists of several successive image patches. Sound channels can also be made up of conversations.

The present invention performs a predetermined process based on a hierarchical structure by using a serial hierarchical dirichlet process (sHDP). That is, the sequential HDP (serial hierarchical dirichlet process) sHDP divides the hierarchical structure of the video stream into lower streams (images and sounds) and processes them separately.

1 is a reference diagram schematically illustrating a data processing schema according to the present invention. Referring to FIG. 1, the present invention may perform estimation using a hidden variable model on an image and perform a separate process from the image on the sound. Unlike the general structure learning approach, the sHDP of the present invention can estimate the story interval by judging the likelihood or the continuity of the data that will be generated by the model trained with each modality (image and sound). The present invention also contemplates estimating the transform point according to the Bayesian methodology. However, in TV dramas, since a sudden change of the image is very frequent, too many change points are estimated when judging the change point using only the likelihood, and the present invention dynamically compensates for the conversion point of each channel. Dynamic Channel Merging can be applied.

1.1 Sequential HDP Structure

In the following, the present invention is described assuming a drama episode as a video stream. The image channel of the drama episode may be expressed as a set of story sections. Theta (ij) is a Visual Word, which can be likened to a guest at the Chinese Restaurant Franchise (CRF). Each restaurant can correspond to a frame, which can be created by measure G (j). These video streams can share a global set of conceptual objects (menu, Phi (k)). One frame may consist of several conceptual objects. The role of the table (t (ij)) is to link the menu with the visual words that are sitting at the table.

2 is a reference diagram illustrating an algorithm for sHDP modeling according to the present invention, and FIG. 3 is a reference diagram illustrating an sHDP model.

In Fig. 2, because of the hierarchical structure, HDP can be a good model for describing the image stream. However, HDP is vulnerable to sudden fluctuations in the image channel, such as camera movement or screen distortion. The Sticky HDP-HMM model was introduced to achieve stable results against these sudden changes. Instead of introducing a new transition probability, the present invention discloses a method of dynamically combining the multi-modalities of a video stream.

HDP models can be used in the image channel. Using this, it is possible to estimate the generation likelihood of the next frame set in the current model when a new frame appears. If the likelihood value is lower than the threshold value, the new HDP model can be estimated using the algorithm of FIG.

However, in the case of the HDP model, only the change in the semantic unit can be detected, and the change in the entire story section cannot be detected. In the case of Ahmed and Xing's paper, the weight of each channel is halved over time, which is fixed and this is not realistic. This is because the most realistic change in the video stream can be found in each channel. Therefore, the present invention further includes a sound channel in the model, which can be represented as shown in FIG.

Equations 1 to 3 are equations for describing the HDP model.

As described above, the present invention can be considered by adding a change of a sound channel to the image model. Here, depending on the recognition performance of the sound channel, the hidden variable model may not be used for the sound channel. The characteristic of a sound channel is that the sound channel may not be exactly the same as the story change, because there may be parts of the scene that are considered to be the same section with no voice. On the other hand, it is less likely that the scene will change during the conversation. Therefore, the present invention discloses a dynamic model which considers the change of both the image channel and the sound channel in view of this point.

1.2 Dynamic Channel Merging

와 대화 감지 측정의 값

을 고려하여 변화 지점을 추정할 수 있다. 이를 수학식으로 표현하면 아래의 수학식 4 내지 수학식 5와 같다.The proposed methodology measures the likelihood change

Values of conversation detection measurements with

The change point can be estimated by considering. This may be expressed as Equation 4 to Equation 5 below.

는 이미지 채널에서의 우도 차이를 표현하고,

는 대화가 계속되는지를 판별할 수 있다.

는

가 주어졌을때

가 나올 우도에 대한 함수이다.

는 현재 프레임의 우도가 이전 프레임의 우도보다 낮을 경우 값이 1이 된다.

는 우도의 급하락 후에 급상승이 새로운 씬이 나타났음을 의미한다는 insight를 반영할 수 있다. 사운드 모델의 경우 대화의 연속성을 모델링 할 수 있으며, 따라서 수학식 4는 대화가 진행되고 있을 때 우도의 변화가 있을 경우 무시하고, 대화와 대화 사이에 우도 변화가 있을 경우에는 이 지점을 변화 후보 지점으로 간주할 수 있다.
From here,

Represents the likelihood difference in the image channel,

Can determine if the conversation continues.

The

Is given

Is a function for likelihood to come out.

The value becomes 1 when the likelihood of the current frame is lower than the likelihood of the previous frame.

Reflects the insight that the spikes after a drop in the likelihood indicate a new scene. For the sound model, we can model the continuity of the dialogue, so Equation 4 ignores the likelihood change when the dialogue is in progress, and uses this point if the likelihood changes between dialogue and dialogue. Can be regarded as.

1.3 Posterior sampling

는 시각 단어의 그룹과 그것의 identifier를 연결시킬 수 있다.
Sampling methods and estimation frameworks can be based on Heinrich's methodology. Hyperparameters in this process can be composed of baseline measure H and concentration gamma. Where Gamma and Alpha represent the current identifier and popularity of the hidden model. We estimate the hyperparameter of DP as shown in the HDP paper. Given x, the change in (k, t) should be estimated for ex post estimation. From here,

Can associate a group of visual words with its identifier.

1.3.1 Sampling an identifier and a conceptual object

In assigning identifiers, the preferred conceptual object is to maintain its popularity. It should be possible to select a new identifier to avoid biasing preferences. Therefore, the present invention can assume the following. (1) People or large objects are more likely to reappear in the same story. (2) A new object is needed to describe the new frame. According to this intuition, the identifier may be sampled in the following Equation 6 and the conceptual object may be sampled in the Equation 7.

는

와 관련된 파라미터(시각 단어)의 갯수를 나타낸다.

는

가 주어 졌을 때

를 재생성할 우도를 나타낸다.

는 이전 스토리의 끝으로부터 현재 scene까지의 시간 경과를 나타낸다.

는 conceptual object t의 pupularity를 나타낸다.
From here,

The

Indicates the number of parameters (visual words) associated with.

The

When was given

Indicates the likelihood to regenerate.

Indicates the time lapse from the end of the previous story to the current scene.

Represents the pupularity of the conceptual object t.

2. Experiments

Hereinafter, the results of applying the video stream analysis method according to the present invention to the division of TV drama episodes will be described. In general, a story is defined as meaning "segment of news that includes two or more homogenous independent clauses describing an event," but in the present invention, for the purpose of semantic distinction of stories, "homogeneous topics" Segment consisting of multiple dialogues and images with " The segment in the present invention is similar to the segment in the general definition in that it splits the stream into homogeneous segments.

2.1 Data and Representation

2.1.1 Data

Nineteen subjects manually recorded the story transformation points in the data. The total length of the test data was 125 minutes and 30 seconds and a total of 7530 scenes were used. Since the probability that many people judge the exact same point in time as a transformation point is very small, the interval of the transformation point is defined as the transformation interval.

Table 1 reflects the statistical characteristics of each episode. In Table 1, the conversion interval of 1 second is fine, but 11 seconds is somewhat longer. This maximum interval corresponds to a scene that shows the rapid change of the sky or city background. In such a scene, it is difficult to say that a specific point is a conversion point.

2.1.2. Representation

Visual words were extracted using SIFT. Scenes of the video stream are composed of images and sound data, and in the case of sound channels, the MFCC (Mel Frequency Cepstral Coefficients) extraction method can be processed. In order to process multi-modal data, other researchers have noted the instantaneous co-occurrence of structures in each modality, but it is difficult to use this method because it is difficult to create a common structure in the matter of story transition point determination. Thus, the present invention uses sHDP to process the image and, in the case of sound, determine the continuity of the conversation.

2.1.3 Experimental Results

Table 2 lists the Precision, recall, and F1 measures. Table 2 presents the baseline true values judged by the person used to compare the performance of the proposed methodology. The number of transform points detected in each episode is 156, 186, and 176, respectively. In addition, there are 43, 39, and 39 conversion intervals, respectively. Since performance depends on the threshold for likelihood, only the best results are listed in the table. 4 to 6 are graphs showing trends of recall, precision and correct answers of episodes 1 to 3, respectively. The number of recall, precision, and correct answers according to the threshold is shown. Equation 8 is an equation for describing recall and precision, and Equation 9 is an equation for describing a threshold.

7 is a graph showing the F1 measure in episodes 1-3. As shown in FIG. 7, the larger the likelihood threshold, the higher the recall and precision, and the lower the correct answer rate. The present invention introduces a model of the sound channel to enable more accurate conversion point determination.

8 is a graph showing the effect of the channel merging method according to the present invention.

In FIG. 8, the solid line represents the likelihood of the image channel and the dotted line represents the signal of the sound channel. The vertical lines are the transform points estimated by the proposed methodology. As shown in FIG. 8, although there are rapid and frequent changes in both channels, it can be determined that there are four transformation points using the proposed methodology.

9 is a block diagram illustrating an embodiment of a video strip analysis apparatus according to the present invention. The video strip analyzing apparatus 100 illustrated in FIG. 9 is a predetermined apparatus configured to perform the above-described contents, and thus descriptions of the same or equivalent contents as those described above will be omitted. The invention will be clearly understood.

Referring to FIG. 9, the video strip analysis apparatus 100 may include a separation module 110, an image learning module 120, an image database 121, a sound learning module 130, a sound database 131, and a control module 140. ) May be included.

The separation module 110 may separate the video stream into an image stream and a sound stream.

The image learning module 120 may extract a visual word for the separated image stream and estimate the likelihood of the current frame set with respect to the following frame set.

The image database 121 may store likelihood data estimated by the image learning module 120.

The sound learning module 130 may model the separated sound stream based on the continuity of the dialogue, and estimate the likelihood of the current frame set with respect to the following frame set by using the generated model.

The sound database 131 may store likelihood data estimated by the sound learning module 130.

The control module 140 checks peak points exceeding a predetermined threshold value based on the likelihood trends estimated by the image learning module 120 and the sound learning module 130, and determines the peak points of the image stream and the sound stream. In association to determine the story transition points of the video stream.

In one embodiment, the control module 140 may use a serial Hierarchical Dirichlet Process (sHDP) that divides and processes an image channel and a sound channel into sub streams, respectively.

In one embodiment, the control module 140 has a second peak point of the likelihood estimated by the sound learning module 130 within a predetermined time range around the first peak point of the likelihood estimated by the image learning module 120. If present, the first peak point may be determined as the story change point.

10 is a flowchart illustrating an embodiment of a video strip analysis method according to the present invention. Since an embodiment of the video strip analysis method disclosed in FIG. 10 is performed by the video strip analysis apparatus of FIG. 9, it may be supported by the above description with reference to FIGS. 1 to 9.

Referring to FIG. 10, the video strip analysis apparatus 100 may classify video stream data into an image channel and a sound channel (step S1010), and configure learning models for each of the divided image channel and the sound channel ( Step S1020).

The video strip analyzing apparatus 100 may estimate the likelihood of the current frame set with respect to the following frame set by using the learning models of the configured image channel and the sound channel (step S1030).

The video strip analyzing apparatus 100 may record the estimated trend of the likelihood and determine a story conversion point of the video stream by using a peak point exceeding a preset threshold value among the recorded likelihood trends (step S1040).

In one embodiment of steps S1010 to S1020, the video strip analysis apparatus 100 separates the video stream data into image channels (image streams) and sound channels (sound streams), and visually identifies the visual words for the separated image channels. The learning model of the image channel can be constructed.

Here, the video strip analysis apparatus 100 may apply a Scale Invariant Feature Transform (SIFT) to the image channel to extract a visual word to construct a learning model of the image channel.

In an embodiment of steps S1010 to S1020, the video strip analysis apparatus 100 may extract a feature using a Mel Frequency Cepstral Coefficients (MFCC) algorithm for the separated sound channel.

In one embodiment, the learning model of the sound channel may perform modeling based on the continuity of the dialogue reflected in the sound channel.

In one embodiment of the step S1040, the video strip analysis apparatus 100 of the likelihood estimated from the learning model of the sound channel within a predetermined time range around the first peak point for the likelihood estimated from the learning model of the image channel. If the second peak point exists, the first peak point can be determined as the story change point.

In one embodiment of step S1040, the video strip analysis apparatus 100 uses the serial Hierarchical Dirichlet Process (sHDP) to divide the image channel and the sound channel into sub streams, respectively, to process the transform points. You can decide.

It will be apparent to those skilled in the art that various modifications and variations can be made in the present invention without departing from the spirit or scope of the present invention as defined by the following claims It can be understood that

Video Strip Analysis Device (100)
Separation Module (110) Image Learning Module (120)
Image Database 121 Sound Learning Module 130
Sound Database 131 Control Module 140

Claims (11)

A video stream analysis method performed in a video strip analysis apparatus capable of receiving a video stream and analyzing a story transformation of the received video stream,
(a) dividing video stream data into an image channel and a sound channel, and constructing learning models for each of the divided image channel and the sound channel;
(b) estimating the likelihood of the current frame set with respect to the following frame set using learning models of the configured image channel and sound channel; And
(c) recording the estimated likelihood trend and determining a story transformation point of the video stream by using a peak point exceeding a preset threshold value among the recorded likelihood trends;
Video stream analysis method comprising a.
The method of claim 1, wherein step (a)
Separating the video stream data into the image channel and the sound channel; And
Constructing a learning model of an image channel using visual words for the separated image channel;
Video stream analysis method comprising the.
The method of claim 2, wherein constructing the learning model of the image channel comprises:
Extracting the visual word by applying a scale invariant feature transform (SIFT) to the image channel;
Video stream analysis method comprising the.
3. The method of claim 2, wherein step (a)
Extracting a feature on the separated sound channel using a Mel Frequency Cepstral Coefficients (MFCC) algorithm;
Video stream analysis method comprising the.
The method of claim 1, wherein the learning model of the sound channel
Modeling based on the continuity of dialogue reflected in the sound channel
Video stream analysis method, characterized in that.
2. The method of claim 1, wherein step (c)
If the second peak point of the likelihood estimated by the learning model of the sound channel is present within a predetermined time range around the first peak point of the likelihood estimated by the learning model of the image channel, the first peak point is described as the story. Determining a conversion point;
Video stream analysis method comprising the.
2. The method of claim 1, wherein step (c)
Determining the conversion point using a sequential hierarchical dirichlet process (SHDP) for dividing and processing the image channel and the sound channel into lower streams, respectively;
Video stream analysis method comprising the.
A video strip analysis apparatus capable of receiving a video stream and analyzing a story transformation of the received video stream,
A separation module for separating the video stream into an image stream and a sound stream;
An image learning module for extracting a visual word for the separated image stream and estimating a likelihood of a current frame set with respect to a following frame set by using the same;
A sound learning module performing modeling on the separated sound streams based on continuity of dialogue and estimating a likelihood of a current frame set with respect to a subsequent frame set using the generated model; And
A peak point exceeding a predetermined threshold value is identified based on a trend of the likelihood estimated by the image learning module and the sound learning module, and the story change point of the video stream is determined by correlating the peak points of the image stream and the sound stream. A control module;
Video strip analysis device comprising a.
9. The apparatus of claim 8, wherein the control module
Determining the conversion point using a sequential hierarchical dirichlet process (sHDP) that divides the image channel and the sound channel into sub streams, respectively.
Video strip analysis device, characterized in that.
9. The apparatus of claim 8, wherein the control module
If the second peak point of the likelihood estimated by the sound learning module is present within a predetermined time range around the first peak point of the likelihood estimated by the image learning module, the first peak point is determined as the story conversion point. that
Video strip analysis device, characterized in that.
A recording medium having recorded thereon a program for executing a video stream analysis method,
The program is a program that can be run in the video strip analysis apparatus that can receive the video stream and analyze the story transformation of the received video stream,
(a) dividing video stream data into an image channel and a sound channel and configuring learning models for each of the divided image channel and sound channel;
(b) estimating the likelihood of the current frame set with respect to the following frame set using learning models of the configured image channel and sound channel; And
(c) recording the estimated likelihood trend and determining a story conversion point of the video stream by using a peak point exceeding a preset threshold value among the recorded likelihood trends;
And a recording medium.

Images