An image data labeling device and its training method, and an image data labeling method thereof.

Abstract

This invention provides an image data labeling device for labeling image data acquired by a vehicle camera. The device includes: a prediction module for predicting the classification task corresponding to the labeling task; a target detection module for performing target detection processing on the image data to detect targets in the image data; a classification execution module for executing the classification task predicted by the prediction module based on the image data processed by the target detection module to obtain a classification task execution result; and a judgment module for determining whether to execute the labeling task on the image data based on the classification task execution result. This invention transforms the labeling task into a simple classification task, enabling complex labeling involving both vision and reasoning, significantly reducing manual labeling input, offering greater flexibility, and greatly saving labeling time.

Description

Technical Field

[0001] This invention relates to the field of image processing, specifically to the field of image data processing in autonomous driving, and more specifically to an image data labeling device and its training method, and an image data labeling method. Background Technology

[0002] With the development and advancement of science and technology, artificial intelligence (AI) technology has become increasingly mature and is being applied to all aspects of human production and life. A crucial characteristic of AI is its ability to learn from the features of existing data to process new data. This necessitates the mention of data labeling tasks, which involve processing existing data and labeling it with tags such as characteristics, objectives, and classifications. These tags then guide AI learning. Therefore, labeling is a common requirement in AI technology. Especially in the field of autonomous driving, labeling collected data such as camera images and radar point cloud files is extremely important.

[0003] Current technologies primarily label image data using methods such as object detection and image classification. While these methods can automate labeling, they still have many limitations. For example, image data captured by cameras in autonomous vehicles requires higher precision in labeling. It's not enough to simply detect the presence of vehicles within the field of view; the positional relationships between the vehicle and other vehicles are also crucial, such as more precise target information like "vehicle ahead," "vehicle to the left front," and "vehicle to the left rear." For instance, object detection algorithms can only detect the presence of vehicles in the image data, and image classification algorithms can only classify images as vehicles. Therefore, to better meet the refined labeling needs of autonomous driving and other similar fields, there is an urgent need for an efficient and flexible labeling solution. Summary of the Invention

[0004] Therefore, the purpose of this invention is to overcome the shortcomings of the prior art and provide an image data labeling device and its training method, as well as an image data labeling method.

[0005] According to a first aspect of the present invention, an image data labeling apparatus is provided for labeling image data acquired by a vehicle camera. The apparatus includes: a prediction module for predicting a classification task corresponding to the labeling task; a target detection module for performing target detection processing on the image data to detect targets in the image data; a classification execution module for executing the classification task predicted by the prediction module based on the image data processed by the target detection module to obtain a classification task execution result; and a judgment module for determining whether to execute the labeling task on the image data based on the classification task execution result.

[0006] Preferably, the marking task includes multiple marking elements, and correspondingly, the classification task includes multiple classification sub-tasks, wherein each classification sub-task corresponds to at least one marking element.

[0007] Preferably, the prediction module includes multiple prediction sub-modules connected in series, wherein each prediction sub-module is used to predict at least one classification sub-task corresponding to a labeled feature.

[0008] Preferably, the marking elements include at least one of the following: reference frame, orientation, conditions, target type, weather, and light intensity.

[0009] Preferably, the classification execution module is used to determine the execution sub-modules corresponding to the plurality of classification sub-tasks respectively, and to assemble the execution sub-modules corresponding to the plurality of classification sub-tasks according to the execution logic.

[0010] Preferably, the prediction module is further configured to select a target prediction module that matches the labeling task from a plurality of candidate prediction modules, and use the target prediction module to predict the classification task corresponding to the labeling task.

[0011] Preferably, the classification execution module is further configured to select a target classification execution sub-module that matches the target prediction module from a plurality of candidate classification execution modules, and to execute the classification task using the target classification execution module based on the image data processed by the target detection module to obtain the classification task execution result.

[0012] According to a second aspect of the present invention, a method for training an apparatus as described in the first aspect of the present invention is provided, the method comprising: S1, acquiring a dataset containing multiple image data collected by cameras in one or more directions of a vehicle, each image data including multiple images and one or more labeling task labels; S2, training the apparatus to convergence using the dataset acquired in step S1, and updating the model parameters in the apparatus using a classification loss.

[0013] According to a third aspect of the present invention, an image data labeling method is provided, the method comprising: acquiring video image data captured by a vehicle camera; formulating a labeling plan and configuring labeling tasks for the labels according to the formulated labeling plan; using the apparatus described in the first aspect of the present invention to perform labeling processing on the video image data according to the labeling tasks, and storing the labeling results of the video image data. Preferably, the method further comprises: modifying the labeling plan in response to the need to add or delete labels, and configuring labeling tasks for the labels in the modified labeling plan.

[0014] Compared with existing technologies, the advantages of this invention are as follows: This invention can utilize vision plus reasoning technology to transform the labeling task into a simple classification task, thereby flexibly realizing the labeling task of complex labels. In existing technologies, typical labeling tasks use models to label images with simple, fixed labels. Therefore, any label update means that the model needs to be updated to adapt to the new label, or manual classification is required. This invention effectively solves the problem of labeling flexibility. For different labeling tasks, the corresponding classification task can be predicted through a prediction module without introducing a new model, significantly reducing the investment in manual labeling. Furthermore, in existing technologies, conventional labeling processes usually require determining all labels before the process begins. Any addition or deletion of labels requires modification of the entire architecture. However, using the image data labeling device of this invention, only the labeling elements need to be added in the step of configuring the labeling problem (i.e., the labeling task), without modifying the overall device structure. For example, for two significantly different labeling tasks—"There is a car behind my car on the left" and "There is a car in front of my car changing lanes to the left"—the prediction and classification execution modules proposed in this invention can adaptively implement both tasks without modifying the entire labeling device, offering greater flexibility and robustness. Furthermore, in existing technologies, labeling tasks are typically performed with one model per label. If complex labels involving multiple labeling elements are involved, labeling a single frame of image data requires sequential processing through multiple models, which is time-consuming. The labeling device proposed in this invention, however, can perform labeling tasks for complex labels on any frame of image data based on visual inference algorithms, significantly reducing labeling time. Attached Figure Description

[0015] The embodiments of the present invention will be further described below with reference to the accompanying drawings, wherein:

[0016] Figure 1 This is a schematic diagram of the frame structure of the image data marking device according to an embodiment of the present invention;

[0017] Figure 2 This is a schematic diagram of the prediction module structure according to an embodiment of the present invention;

[0018] Figure 3 This is a schematic diagram illustrating an example of a prediction module according to an embodiment of the present invention;

[0019] Figure 4 This is a schematic diagram illustrating the working principle of the prediction module, classification module, and target detection module according to an embodiment of the present invention.

[0020] Figure 5 This is a schematic diagram illustrating the working principle of an image data marking device according to an embodiment of the present invention.

[0021] Figure 6This is a schematic diagram of the image data marking method according to an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention is further described in detail below through specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0023] As described in the background section, existing image labeling methods cannot achieve refined labeling tasks that combine vision and reasoning. To address this issue, this invention proposes a scheme for constructing an image data labeling device through labeling task prediction and end-to-end visual reasoning classification.

[0024] According to one embodiment of the present invention, such as Figure 1 As shown, an image data labeling device of the present invention includes: a prediction module, a target detection module, a classification execution module, and a judgment module. The prediction module is used to predict the classification task corresponding to the labeling task; the target detection module is used to perform target detection processing on the image data to detect targets in the image data; the classification execution module is used to execute the classification task predicted by the prediction module based on the image data processed by the target detection module to obtain a classification task execution result; and the judgment module is used to determine whether to execute the labeling task on the image data based on the classification task execution result.

[0025] To better understand this invention, each module will be described in detail below.

[0026] I. Prediction Module

[0027] Existing labeling schemes can only perform simple labeling tasks on image data and cannot complete complex labeling tasks involving both vision and reasoning. Therefore, this invention proposes a scheme that transforms the complex labeling task involving both vision and reasoning into a simple classification task through a prediction module, then executes the classification task, and determines whether a labeling task needs to be performed based on the classification result.

[0028] According to an embodiment of the present invention, the labeling task includes multiple labeling elements, and correspondingly, the classification task includes multiple classification sub-tasks, wherein each classification sub-task corresponds to at least one labeling element. The labeling elements may include independent labeling objects decomposed from the labeling task. For example, in the field of autonomous driving, the labeling elements involved may include location, weather, reference objects, target type, etc. In a specific example, a labeling task is "There is a car to the left rear of the vehicle," that is, labeling the image data of "There is a car to the left rear of the vehicle." The labeling elements of this labeling task include "reference frame (vehicle)," "location (left rear)," "condition (present)," and "target type (car)." Existing technologies cannot perform such complex labeling tasks, while the solution based on the present invention can predict these labeling elements into different classification sub-tasks through a prediction module. According to an embodiment of the present invention, as... Figure 2 As shown, the prediction module includes multiple serially connected prediction sub-modules (represented by prediction sub-module 1, prediction sub-module 2, ..., prediction sub-module N), wherein each prediction sub-module is used to predict at least one classification sub-task corresponding to a labeled element. According to one embodiment of the present invention, the prediction module is a sequence-to-sequence neural network model trained based on natural language processing. By training on historical labeled task sentences, the accuracy of the prediction module in judging the classification task corresponding to the labeled task can be improved, enabling the prediction module to analyze various classification sub-tasks, including but not limited to subject, location, target type, weather, and light. Since model training based on natural language processing is a technique known to those skilled in the art, it will not be elaborated upon here. The prediction sub-module can be GRU, LSTM, etc., and is not specifically limited here. However, for ease of description, GRU will be used as the prediction sub-module in subsequent embodiments of the present invention. It should be understood that this does not mean that the prediction sub-module can only be a GRU model.

[0029] According to one embodiment of the present invention, the prediction module is further configured to select a target prediction module that matches the labeling task from a plurality of candidate prediction modules, and use the target prediction module to predict the classification task corresponding to the labeling task. The candidate prediction modules corresponding to different labeling tasks are all based on pre-trained sequence-to-sequence neural network models using natural language processing. The prediction module selects the model that matches the labeling task based on the labeling task. Different candidate prediction modules can be pre-set for different combinations of labeling elements and / or task logic. For example, for combinations of labeling elements, candidate prediction model 1 can be used for labeling tasks containing labeling elements such as orientation, reference frame, conditions, and target object type; candidate prediction model 2 can be used for labeling tasks containing labeling elements such as target type, weather, and light intensity. Task logic refers to the logical relationship between tagging elements, which may include chain structure and tree structure. Candidate prediction model 3 can be used for tagging tasks with chain structure, such as tagging task "There is a car behind the left of the car". Candidate prediction model 4 can be used for tagging tasks with tree structure, such as tagging task "There is a car behind the left of the car and there is no truck in front of the car". According to an example of the present invention, taking the labeling task "There is a car behind the left rear of the vehicle" as an example, the labeling elements include "reference frame (vehicle)," "location (left rear)," "condition (present)," and "target type (car)." The classification subtask corresponding to the labeling element "reference frame" is to classify the image data according to whether the reference frame belongs to the vehicle. Similarly, the classification subtask corresponding to the labeling element "location" is to classify the image data according to whether it is image data from the left rear of the vehicle (including left and rear). The classification subtask corresponding to the labeling element "condition" is to classify the image data according to whether it contains a car. The subtask corresponding to the labeling element "target type" is to classify the image data according to whether it contains a car. Figure 3 As shown in the example, the prediction module selects the target prediction module that matches "there is a car behind the left rear of the vehicle" to predict the classification subtask corresponding to this labeling task. Thus, the prediction module can transform the complex labeling task involving both vision and reasoning into a simple classification task, facilitating the subsequent execution of the classification task by the subsequent module. It should be noted that when there are multiple labeling tasks, the target prediction module corresponding to each labeling task can be selected to predict its corresponding classification task.

[0030] For other labeling tasks, such as "It's raining in front of my car" or "There's a car changing lanes to the left in front of my car", you can refer to the above examples and select the target prediction module corresponding to the labeling task to perform the prediction of the classification sub-task. We will not go into further detail here.

[0031] II. Target Detection Module

[0032] As described above regarding the prediction module, after converting the labeling task into a classification task, the most crucial aspect of this invention is classifying the targets in the image data. Therefore, this invention employs a target detection module to pre-detect targets in the image data to facilitate subsequent classification processing. The target detection module can utilize common target detection models such as R-CNN, Fast R-CNN, YOLO, and SSD; this invention does not impose any specific limitations.

[0033] III. Classification Execution Module

[0034] The classification execution module is used to execute the classification task corresponding to the labeling task predicted by the prediction module. As described in the previous embodiments, the visual plus reasoning labeling task generally corresponds to multiple classification sub-tasks. The classification execution module determines the execution sub-module corresponding to each classification sub-task and assembles the execution sub-modules corresponding to the multiple classification sub-tasks according to the execution logic. For example, taking the previous example "there is a car behind the left of the car" as an example, the corresponding classification sub-tasks include: classifying the image data according to whether the reference frame belongs to the car, classifying the image data according to whether it is the image data behind the left of the car (including left and rear), classifying the image data according to whether it contains a car, and classifying the image data according to whether it contains a car. For ease of description, the above classification sub-tasks are referred to as classification sub-task 1, classification sub-task 2, classification sub-task 3, and classification sub-task 4, respectively. For all classification subtasks, they need to be executed according to certain execution logic to obtain accurate results. Correspondingly, the execution sub-modules need to be assembled according to this logic. For the classification subtasks in the example above, the execution logic is to execute classification subtask 1, classification subtask 2, classification subtask 3, and classification subtask 4 in sequence. That is, first, the image data is classified according to whether the reference frame belongs to the vehicle. Then, the image data with the reference frame as the vehicle is classified according to whether it is the image data of the vehicle's left rear (including left and rear). Next, the image data of the vehicle's left rear is classified according to whether there is a car in the vehicle's left rear. Finally, the image data of the vehicle with a car in the vehicle's left rear is classified according to whether there is a sedan in the vehicle's left rear. Finally, the classification task execution result is obtained.

[0035] According to an embodiment of the present invention, the classification execution module is further configured to select a target classification execution module that matches the target prediction module from a plurality of candidate classification execution modules, and to perform the classification task using the target classification execution module based on the image data processed by the target detection module to obtain the classification task execution result. For different candidate classification execution modules, different candidate prediction models can be pre-set. For example, taking the candidate prediction models mentioned in the previous embodiment as an example, candidate classification execution module 1 can be used to perform a classification task corresponding to labeling elements such as orientation, reference frame, condition, and target object type for candidate prediction model 1; candidate classification execution module 2 can be used to perform a classification task corresponding to labeling elements such as orientation, reference frame, condition, and target object type for candidate prediction model 2; candidate classification execution module 3 can be used to perform a classification task corresponding to a labeling task such as "There is a car behind the left rear of the vehicle" for candidate prediction model 3; and candidate classification execution module 4 can be used to perform a classification task corresponding to a labeling task such as "There is a car behind the left rear of the vehicle and no truck in front of the vehicle" for candidate prediction model 4.

[0036] According to one example of the present invention, taking a vehicle that includes a front-side camera, a rear-side camera, and a left-side camera as an example, the image data collected by these three cameras is labeled with the message "There is a car behind the left rear of the vehicle." Figure 4 As shown, the prediction module obtains the classification tasks (sub-tasks 1, 2, 3, and 4) corresponding to the labeling task "There is a car behind the left side of the vehicle" and sends them to the classification execution module. The target detection module performs target detection on the data collected by the three cameras and then sends it to the classification execution module for classification task execution. The classification execution module executes each sub-task in the order of sub-task 1, sub-task 2, sub-task 3, and sub-task 4 according to the execution logic of the classification task. First, sub-task 1 is executed, which involves filtering image data with the vehicle as the reference frame. Since the image data in this example are all image data with the vehicle as the reference frame, therefore... Figure 4 The execution process of subtask 1 is not shown; only subtasks 2, 3, and 4 are shown. Figure 4 As shown, classification subtask 2 categorizes the image data corresponding to the vehicle according to whether it is in the left rear to filter out the image data corresponding to the left and right cameras. Then, classification subtasks 3 and 4 are performed on the image data corresponding to the left and right cameras. That is, the data is classified according to whether there is a car in the left rear to filter out the image data with a car in the left rear. Then, the image data with a car in the left rear is classified with cars as the target to filter out the image data with cars in the left rear. Finally, the execution result of the classification task is obtained.

[0037] For other labeling tasks, such as "It's raining in front of my car" or "There's a car changing lanes to the left in front of my car", you can refer to the above examples and select the classification execution module that matches the target prediction module corresponding to the labeling task to perform the classification task. We will not go into further detail here.

[0038] Similarly, as described above regarding the prediction module, when there are multiple labeling tasks, the classification tasks corresponding to each labeling task can be executed in the same manner as in the above embodiments to obtain the classification execution results of multiple labeling tasks. This will not be elaborated further here.

[0039] IV. Judgment Module

[0040] The judgment module is used to determine whether to perform the labeling task on the image data based on the result of the classification task. For example, still using... Figure 4 For example, as described in the example, Figure 5 As shown, the judgment module determines whether to perform the labeling task "There is a car behind the left side of the vehicle" on the image data based on the execution result of the classification execution module. If the classification execution module selects image data with "There is a car behind the left side of the vehicle" in the classification result after performing the classification task, then the labeling task is performed on the selected image data.

[0041] According to an embodiment of the present invention, the judgment module can use a softmax classifier to classify and judge the labeling task based on the classification execution result of the classification execution module, and the output is 0 or 1, where 0 represents that no image data was filtered out in the classification task execution result, i.e., the labeling task "does not exist", and 1 represents that image data was filtered out in the classification task execution result, i.e., the labeling task "exists".

[0042] Furthermore, it should be noted that the labeling device of this invention includes multiple types of neural networks. To improve the accuracy of the device, this invention also proposes a supervised learning-based model training method, which uses existing historical data as the training set to supervise the training of the image data labeling device. Specifically, for the prediction module, historical labeling task statements can be used as input, and the corresponding classification tasks can be used as output for training. Different candidate prediction modules can be trained for different labeling tasks. For the target detection module, historical image data collected by a vehicle camera can be used as input, and the targets corresponding to the historical image data can be used as output for training. For the classification execution module and the judgment module, they can be trained as a whole, using historical image data as input and the corresponding labeling tasks as output. Alternatively, they can be trained separately, and different candidate classification execution modules can be trained for different labeling tasks. Combining the trained modules forms the image data labeling device. Since supervised training is a technique known to those skilled in the art, this invention will not elaborate further.

[0043] When the image data marking apparatus of the present invention is used to perform image data marking tasks, in general, as follows: Figure 6 As shown, the process includes: acquiring data acquisition files, specifying a labeling plan, configuring labeling tasks, calling the visual inference model (the image data labeling device of this invention) for labeling, and storing the labeling results. Acquiring data acquisition files refers to acquiring video image data captured by a vehicle camera; specifying a labeling plan refers to defining which labels need to be applied; configuring labeling tasks refers to configuring labeling tasks for the labels according to the specified labeling plan; calling the visual inference model for labeling refers to using the image data labeling device of this invention to execute the configured labeling tasks on the data acquisition files. When there is a need to add or delete labels, it is only necessary to modify the labeling plan and configure the labeling tasks based on the modified labeling plan.

[0044] As demonstrated by the above embodiments, this invention can utilize vision plus reasoning technology to transform the labeling task into a simple classification task, thereby flexibly realizing the labeling of complex labels. In the prior art, typical labeling tasks involve using models to label images with simple, fixed labels. Therefore, any label update requires updating the model to adapt to the new label or manual classification. This invention effectively solves the problem of labeling flexibility. For different labeling tasks, the prediction module can predict the corresponding classification task without introducing a new model, significantly reducing the need for manual labeling. Furthermore, in the prior art, conventional labeling processes typically require determining all labels before the process begins. Adding or deleting labels requires modifying the entire architecture. However, using the image data labeling device of this invention, only the labeling elements need to be added in the step of configuring the labeling problem (i.e., the labeling task), without modifying the overall device structure. For example, for two significantly different labeling tasks—"There is a car behind and to the left of my car" and "There is a car in front of my car changing lanes to the left"—the prediction and classification execution modules proposed in this invention can adaptively implement both tasks without modifying the entire labeling device, offering greater flexibility and robustness. Furthermore, in existing technologies, labeling tasks typically involve one model per label. If complex labels with multiple labeling elements are involved, labeling a single frame of image data requires sequential processing through multiple models, which is time-consuming. The labeling device proposed in this invention, however, can perform labeling tasks for complex labels on any frame of image data based on visual inference algorithms, significantly reducing labeling time.

[0045] It should be noted that although the steps are described in a specific order above, it does not mean that the steps must be executed in the above specific order. In fact, some of these steps can be executed concurrently or even in a different order, as long as the required function can be achieved.

[0046] This invention can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of the invention.

[0047] Computer-readable storage media can be tangible devices that hold and store instructions for use by an instruction execution device. Computer-readable storage media can be, for example, including but not limited to, electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination thereof.

[0048] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. An image data marking device for marking image data acquired by a vehicle camera, characterized in that, The device includes: A prediction module is used to predict the classification task corresponding to the labeling task; wherein, the labeling task includes multiple labeling elements, and correspondingly, the classification task includes multiple classification sub-tasks, each classification sub-task corresponding to at least one labeling element; the labeling elements include at least one of the following: reference frame, orientation, condition, target type, weather, and light intensity; the prediction module includes multiple serially connected prediction sub-modules, wherein each prediction sub-module is used to predict the classification sub-task corresponding to at least one labeling element; The target detection module is used to perform target detection processing on the image data in order to detect targets in the image data; The classification execution module is used to execute the classification task predicted by the prediction module based on the image data processed by the target detection module to obtain the classification task execution result; wherein, the classification execution module is used to determine the execution sub-modules corresponding to the plurality of classification sub-tasks respectively, and assemble the execution sub-modules corresponding to the plurality of classification sub-tasks according to the execution logic; The judgment module is used to determine whether to perform the labeling task on the image data based on the result of the classification task.

2. The apparatus according to claim 1, characterized in that, The prediction module is further configured to select a target prediction module that matches the labeling task from multiple candidate prediction modules, and use the target prediction module to predict the classification task corresponding to the labeling task.

3. The apparatus according to claim 2, characterized in that, The classification execution module is further configured to select a target classification execution module that matches the target prediction module from multiple candidate classification execution modules, and to execute the classification task using the target classification execution module based on the image data processed by the target detection module to obtain the classification task execution result.

4. A method for training the apparatus as described in any one of claims 1-3, characterized in that, The method includes: S1. Obtain a dataset containing multiple image data collected by cameras in one or more directions of the vehicle. Each image data includes multiple images and one or more labeling task labels. S2. The device is trained to convergence using the dataset obtained in step S1, and the parameters of each model in the device are updated using classification loss.

5. An image data labeling method, characterized in that, The method includes: Acquire video image data captured by the vehicle's camera; Develop a labeling plan and configure labeling tasks for the labels according to the developed plan; The device described in any one of claims 1-3 is used to perform marking processing on video image data according to the marking task, and the marking results of the video image data are stored.

6. The method according to claim 5, characterized in that, The method further includes: In response to the need to add or delete tags, modify the tag plan and configure tagging tasks for the tags in the modified tag plan.

7. A computer-readable storage medium, characterized in that, It stores a computer program that can be executed by a processor to implement the steps of the method as described in any one of claims 4-6.

8. An electronic device, characterized in that, include: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the electronic device to perform the steps of the method as described in any one of claims 4-6.

Images