Intelligent image analysis platform based on automatic data training for each situation

Abstract

The image analysis system includes a plurality of processes independently executed in a computing device including a central processing unit and a memory. The plurality of processes may include: an image analysis process of storing, for each channel, image data in which one or both of an object and an event are detected using a detection model provided for each channel; metadata generated in association with one or both of the detected object and event, and feedback on the image data and the metadata, wherein the feedback includes normal detection and error detection; and a model advancement process of collecting the image data, the metadata, and the feedback, monitoring the collected feedback, and generating an advanced detection model when drift is detected.

Description

Intelligent video analysis platform based on situational data automatic learning

[0001] The present invention relates to image analysis.

[0002] As social crime increases day by day and science and technology advance, surveillance systems utilizing cameras like CCTV are rapidly improving for crime prevention and monitoring of traffic congestion. Recently, surveillance systems are evolving beyond simple surveillance to intelligent notification systems. Furthermore, with the urgent need for big data collection for efficient monitoring of situations like traffic congestion, it's no exaggeration to say that camera video technology is in a state of flux.

[0003] When a video analytics system is installed in a site requiring high security, external access and maintenance can be difficult. Upgrading the AI ​​model to suit the installation site requires utilizing training data collected on-site. Because collecting training data is practically impossible in sites with limited external access, AI model upgrades can only be accomplished by the users operating the video analytics system on-site. However, these users typically lack the knowledge or development capabilities required for video analytics.

[0004] According to an embodiment of the present invention, an intelligent image analysis system based on context-dependent data automatic learning is provided. The image analysis system may include a plurality of processes independently executed in a computing device including a central processing unit and a memory, wherein the plurality of processes may include an image analysis process that stores, by channel, image data in which one or both of an object and an event are detected using a detection model provided for each channel, metadata generated in association with one or both of the detected objects and events, and feedback on the image data and the metadata, wherein the feedback includes normal detection and error detection, and a model enhancement process that collects the image data, the metadata, and the feedback, and monitors the collected feedback to detect drift, and generates an advanced detection model.

[0005] In one embodiment, the model enhancement process may include a monitoring process for monitoring feedback of the image analysis by channel to initiate model enhancement of a detection model of a channel in which the drift is detected, a labeling process for generating labeled training data including a detection image in which one or both of an object and an event are detected in the image data and indicated by a bounding box and metadata associated with the detection image, a detection model generation process for generating an enhanced detection model by training a base model with the labeled training data, a model verification process for verifying performance of the enhanced detection model, and a model deployment process for deploying the verified detection model to the image analysis.

[0006] In one embodiment, the labeled training data may include first labeled training data used in training a detection model of the channel in which the drift was detected and unused second labeled training data.

[0007] In one embodiment, the model validation process may commonly apply a validation data set selected from the second labeled training data to the detection model of the channel in which the drift was detected and the advanced detection model.

[0008] In one embodiment, the detection model generation process trains two or more base models with different computing power requirements using the same labeled learning data, and a detection model requiring relatively high computing power can be loaded into the labeling process, and a detection model requiring relatively low computing power can be loaded into the image analysis process.

[0009] In one embodiment, the image analysis system may further include a client process that provides feedback on the image data and the metadata stored in the image analysis process.

[0010] In one embodiment, the image analysis process uses detection models learned for multiple channels, and only the detection model used for the channel where the drift is detected can be advanced.

[0011] The present invention is described below with reference to embodiments illustrated in the accompanying drawings. To facilitate understanding, identical components are assigned the same reference numerals throughout the accompanying drawings. The components depicted in the accompanying drawings are merely exemplary embodiments implemented to illustrate the present invention and are not intended to limit the scope of the present invention. In particular, the accompanying drawings may slightly exaggerate some of the elements depicted in the drawings to facilitate understanding of the invention.

[0012] Figure 1 is a schematic diagram illustrating an intelligent image analysis system based on situation-specific data automatic learning.

[0013] Figure 2 is a diagram functionally illustrating the image analysis process.

[0014] Figure 3 is a diagram functionally illustrating the model enhancement process.

[0015] Figure 4 is a diagram illustrating an example of a process for automatically triggering a model enhancement process by data drift.

[0016] Figures 5 and 6 are diagrams illustrating exemplary scenarios for deploying advanced detection models.

[0017] The present invention is susceptible to various modifications and embodiments. Specific embodiments are illustrated in the drawings and described in detail. However, this is not intended to limit the present invention to specific embodiments, but rather to encompass all modifications, equivalents, and alternatives falling within the spirit and technical scope of the present invention. In particular, the functions, features, and embodiments described below with reference to the accompanying drawings may be implemented alone or in combination with other embodiments. Therefore, it should be noted that the scope of the present invention is not limited to the forms illustrated in the accompanying drawings.

[0018] Throughout the attached drawings, identical or similar elements are referenced using the same drawing reference numerals. For convenience and understanding, the drawings are somewhat exaggerated.

[0019] Figure 1 is a schematic diagram illustrating an intelligent image analysis system based on situation-specific data automatic learning.

[0020] The intelligent image analysis system (10) based on situational data automatic learning can detect changes in the installation site by capturing new data patterns, and can enhance the detection model to maintain detection accuracy. For example, the performance of an image analysis system (10) that is installed in a site requiring a high level of security and cannot collect learning data can also be optimized. Referring to FIG. 1, the image analysis system (10) includes an image analysis system management process (100), an image analysis process (200), and a model enhancement process (300), and may further include an image distribution process (400). The image analysis system management process (100), the image analysis process (200), the model enhancement process (300), and the image distribution process (400) may be processes that are physically executed independently on one computing device, or each executed on two or more computing devices connected through a communication network. Meanwhile, the image analysis system management process (100), the image analysis process (200), the model enhancement process (300), and the image distribution process (400) may include multiple sub-processes operating within them. The computing device includes physical components such as one or more computing devices (e.g., CPU, GPU, AP, etc.), memory, and communication modem chips.

[0021] The video analysis system management process (100), the video analysis process (200), the model advancement process (300), and the video distribution process (400) can exchange data with each other through one or more communication protocols. The communication protocols vary depending on whether they are synchronized, the transmission target, etc., and more than one communication protocol may be used between the processes. For example, the video data provided by the camera (500) is transmitted between the video analysis system management process (100), the video analysis process (200), and the model advancement process (300) through RTSP (Real time streaming protocol), and the metadata generated by the video analysis process (200) can be transmitted between the video analysis system management process (100) and the model advancement process (300) through any one of a plurality of protocols such as REST (Representational State Transfer) API, ZMQ (Zero-Message queue), MQTT (MQ Telemetry Transport), TCP / IP, and Serial.

[0022] Data exchange between processes running on physically independent computing devices can occur via a communications network. The communications network can be a wired, wireless, or hybrid data communication network capable of transmitting data. Wired networks can be dedicated lines or cable networks that support communication protocols for transmitting digital data in packet form. Wireless networks can be communication systems that transmit data using wireless signals, such as CDMA, WCDMA, GSM, EPC, LTE, WiBro, or even Bluetooth and ZigBee, in addition to Wi-Fi.

[0023] The video analysis system management process (100) can communicate with the video analysis process (200) to set up or monitor channels and events, and can communicate with the model enhancement process (300) to manage model enhancement and distribution. In addition, the video analysis system management process (100) can receive video data from the video analysis process (200) and / or the video distribution process (400). By monitoring, the video analysis system management process (100) detects error detection in the video analysis process (200) and records the error detection so that it can be utilized in the model enhancement process (300).

[0024] The image analysis process (200) analyzes image data received from the camera (500). The image analysis process (200) may receive image data directly from the camera (500) or from the image distribution process (400). The camera (500) may be a CCTV, IP camera, infrared camera, etc. that provides image data such as two-dimensional images, three-dimensional images, infrared images, depth images, etc. In addition, the camera (500) includes not only a fixed camera installed in a specific location, but also a mobile camera mounted on a drone or robot.

[0025] The model enhancement process (300) enhances an artificial intelligence model and / or machine learning model (hereinafter referred to as a detection model) used for object and / or event detection in image analysis, and distributes the enhanced detection model to the image analysis process (200). Raw learning data is generated by the image analysis process (200), and the model enhancement process (300) collects the raw learning data and uses it to enhance the detection model.

[0026] The image distribution process (400) is a process applied in a large-scale site, transcoding image data received from a camera (500) and distributing it to the image analysis system management process (100) and / or the image analysis process (200).

[0027] Figure 2 is a diagram functionally illustrating the image analysis process.

[0028] The image analysis process (200) is a process that independently performs the function of analyzing image data to detect events and generating metadata related to the detected events. Referring to FIG. 2, the image analysis process (200) includes a plurality of processes for analyzing image data to generate metadata. The image analysis process (200) basically includes an image data processing module (210) that processes the received camera images to be suitable for image analysis and outputs image data, an object detection module (220) that detects objects in the image data, an object tracking module (230) that tracks the detected objects, an event detection module (240) that detects events related to the objects being tracked and generates metadata, and a detection result storage unit (250). When installed in a large-scale site and including an image distribution process (400), the image data processing module (210) may be omitted. In the structure illustrated in the drawing, the object detection module (220), the object tracking module (230), and the event detection module (240) are shown as separate processes, but it should be understood that object detection, object identification, and object tracking can also be performed as a series of inseparable processes.

[0029] The image analysis process (200) can be equipped with a detection model. The detection model can be a base model of the CNN series that can perform deep learning-based transfer learning, such as ResNet, Vgg, etc. Among the deep learning-based artificial intelligence models, a model that can perform object location and classification through regression analysis is representative of R-CNN (Regional-based Convolutional Neural Network), and Fast-RCNN and Faster-RCNN can be additionally or supplementarily applied to improve speed. The detection model can be executed in one or more of the object detection module (220), the object tracking module (230), and the event detection module (240). The detection model equipped in the image analysis process (200) can be updated by the model enhancement process (300).

[0030] The object detection module (220) detects one or more objects from image data and can classify the detected objects by type. The object detection module (220) may be, for example, an object detection module trained using multiple object images, or an object detection module using a template representing an object.

[0031] The object tracking module (230) tracks the movement of the detected object. The object tracking module (230) can track the object by predicting the movement of the object or by comparing the properties of the object. The object tracking method predicts the movement or detects the properties of the object detected in the first image data (or frame) and compares it with the object detected in the second image data (or frame) to determine whether the two objects are the same object. Two or more object tracking methods can be applied, such as a method of tracking the movement of the object by comparing the macroblock units that constitute the image or a method of detecting and tracking the movement of the object by removing the background from the image so that only the object remains in the image.

[0032] The object tracking module (230) can identify the same object in image data. Multiple cameras (105) can capture objects moving within a region of interest. Accordingly, an object detected at a first point in time may disappear from the camera image at a second point in time and reappear in the image at a third point in time. In one embodiment, the object detection module (220) can detect the same object displayed in two or more images by referring to the spatial relationship between the captured areas at the first to third points in time. In another embodiment, the object detection module (220) can extract attributes of the detected object through deep learning analysis and compare the attributes of the extracted objects to determine whether objects displayed at different points in time are the same object. Depending on the determination result, for example, the probability that two objects match each other (hereinafter, similarity) may be displayed, or an object identifier assigned as a result of image analysis at the first point in time may be reassigned to the object detected in the camera image at the third point in time.

[0033] The object tracking module (230) can utilize appropriate attributes to identify the object, depending on the object type. In the case of a human object, the object tracking module (230) can re-recognize the human object using, for example, visual features, facial features, and gait features. Visual features may include the height of the human object, the type or color of clothing worn, etc.

[0034] The event detection module (240) detects an event that satisfies an event condition based on one or a combination of a detected object, the movement of the object, and the surrounding environment of the object. The event detection module (240) is provided for each channel, and each event detection module (240) can detect an event from the image data of the assigned channel. Meanwhile, the event detection module (240) can be driven by two or more different image analysis methods. For example, a deep learning-based event detection unit extracts attributes of a detected object and determines whether it satisfies a set event condition based on the attributes, and a rule-based event detection unit detects an event that satisfies the event condition based on the movement of the object in an area of ​​interest. Meanwhile, an object that satisfies a first event condition through deep learning analysis can be passed to a rule-based analysis, and it can be determined whether it satisfies a second event condition. If there are multiple event conditions set to be performed through deep learning, there can also be multiple deep learning-based event detection units that operate according to each event condition. Likewise, if there are multiple event conditions set to perform a rule-based method, there may also be multiple rule-based event detection units that operate according to each event condition.

[0035] Event Classification Description Heterogeneous Object Classification Classifies and integrates the attribute values ​​of different types of objects detected by cameras Event Detection Illegal Parking Detects when a car is stopped in a place where it should not be parked and is delayed for a certain period of time Speed ​​and Location Displays the speed and location of the car when the car is driven based on the point where the camera is installed Traffic Volume and Congestion Detects the density and counting of cars on the road Illegal U-turns and reverse driving Detects the direction of cars on the road to detect illegal U-turns and reverse driving Left Turn Sensitivity Displays a left turn signal in conjunction with the signal when a car waiting for a left turn signal has been waiting for a certain period of time People Counting and Congestion Counts the number of people entering and exiting a specific area and detects congestion due to the entry of a certain number of people or more Trespassing Crossing Detects the moment an unauthorized person enters a no-entry zone, trespassing, or trespassing Staying Determines that a detected object is staying when it continues to be detected in the same place for a certain period of time Loitering Detects suspicious behavior such as loitering or lingering for a certain period of time without a specific event after detecting a person object Abandonment New object separated from the detected person object Fall DetectionFalling is detected when the head touches the floor due to an action of falling to the floor while walking or standing. Facial Recognition and DetectionFunction to recognize a person's face in a camera image and extract facial feature points (FFV).Similarity AnalysisCompare the similarity of the facial feature points detected in the camera image with the face DB stored in the DB to detect authorized / unauthorized persons.FightingFunction to measure the fighting posture after detecting a person and analyze multiple actions to detect them.Crowd DensityFunction to determine the density of the area of ​​interest of detected people.Flames / SmokeFunction to determine whether it is a fire by detecting flames and smoke.Elderly and BlindFunction to determine whether an elderly person is in a weak position by detecting a cane or wheelchair.WeaponsFunction to determine whether a person is dangerous by detecting a person and a weapon.ArmsFunction to determine whether a person is wearing armor or not.

[0036] When an event is detected, the event detection module (240) generates metadata including a timestamp associated with the event detection time. The timestamp indicates the event detection time based on the video distribution time. For example, the video provided by the camera (500) is distributed to the video analysis process (200) and the video analysis system management process (100) by the video distribution process (400), and the metadata generated by the video analysis process (200) is subsequently provided to the video analysis system management process (100). Therefore, in the video analysis system management process (100), the synchronization between the video data and the metadata does not match. Therefore, the video analysis system management process (100) can synchronize the video data and the metadata using the timestamp included in the metadata. The detection result storage unit (250) stores the video data in which one or both of the objects and events are detected, and the metadata associated with one or both of the detected objects and events. Meanwhile, the video data and the metadata stored in the detection result storage unit (250) can be classified into either normal detection or error detection through feedback. Raw training data may include image data, metadata, and feedback.

[0037] Figure 3 is a diagram functionally illustrating a model enhancement process, and Figure 4 is a diagram exemplarily illustrating a process of automatically triggering a model enhancement process by data drift.

[0038] Referring to FIGS. 3 and 4 together, the model enhancement process (300) includes a learning database (310), a labeling process (320), a detection model generation process (330), a model storage (340), a model deployment process (350), a client process (360), and a monitoring process (370).

[0039] The learning database (310) collects and stores raw learning data stored by channel in the detection result storage unit (250) of the image analysis process (200), i.e., image data in which one or both of objects and events are detected, and metadata and feedback associated with one or both of the detected objects and events. For example, the learning database (310) may collect and store image data and metadata in which the feedback is error detection. As another example, the learning database (310) may collect and store image data and metadata in which the feedback is normal detection and error detection. Meanwhile, the learning database (310) stores labeled learning data.

[0040] The labeling process (320) detects one or both objects and events from image data stored in the learning database (310) to generate labeled learning data including a detected image and metadata associated with the detected image. The labeling process (320) detects one or both objects and events from the image data to display a bounding box on the detected image, and includes a detection model that classifies the object / event displayed by the bounding box. Compared to the detection model loaded in the image analysis process (200), the detection model loaded in the labeling process (320) has relatively high precision and recall. The detection model may be a detection model corresponding to a channel in which model advancement has been initiated among the detection models registered in the model database (340). The labeling process (320) may be implemented, for example, with a CVAT (Computer Vision Annotation Tool). The labeled learning data may be reviewed by the client process (360).

[0041] The detection model generation process (330) generates an advanced detection model for each channel using labeled training data. The detection model generation process (330) includes data preparation (331), model training (332), and model validation (333). The detection model generation process (330) can be implemented using AirFlow, for example.

[0042] Data preparation (331) obtains learning data labeled by channel from the learning database (310).

[0043] Model training (332) trains a base model with labeled training data to generate an advanced detection model. The labeled training data includes first labeled training data used to train the existing detection model and second labeled training data not used to train the existing detection model. As described above, the second labeled training data is generated to customize the detection model so that it can detect new data patterns through feedback. Therefore, model training (332) can train two or more base models using both the first labeled training data and the second labeled training data. For example, the image analysis process (200) and the labeling process (320) can be run on hardware with different computing power. Therefore, a detection model generated by training two or more base models with different computing power requirements with the same labeled training data can be distributed according to the required performance.

[0044] Model validation (333) verifies the performance of the advanced detection model. The validation data set used for validation is selected from the second labeled training data and can be applied to both the existing detection model and the advanced detection module.

[0045] The performance of the advanced detection module can be verified by one or a combination of Mean Average Precision (mAP), precision, and recall.

[0046] mAP is a metric used to evaluate the performance of a detection model. It represents the average precision when the intersection over union (IoU) threshold for an object is 0.5. mAP measures accuracy by calculating the IoU between the bounding boxes predicted by the detection model and the bounding boxes of the actual objects. mAP is calculated individually for multiple classes and then averaged to measure the overall performance of the detection model. A higher mAP indicates a detection model's ability to detect and classify objects with high accuracy.

[0047] Precision represents the ability of a detection model to identify only relevant objects. For example, if a detection model detects 10 objects, of which 7 are positive, the precision is 0.7.

[0048] Recall represents the detection model's ability to find all relevant instances (all correct bounding boxes). For example, if there are nine correct objects to detect and the detection model correctly identifies six of them, recall is 0.67.

[0049] The model registry (340) registers advanced detection models that have been verified. The model registry (340) is a platform for managing the life cycle of detection models, and manages not only advanced detection models but also detection models created prior to advancement.

[0050] The model deployment process (350) deploys a newly registered advanced detection model in the model registry (340) to the image analysis process (200). Additionally or optionally, the model deployment process (350) may generate a detection engine for the advanced detection model. The detection engine is a detection model optimized for the application environment, e.g., the image analysis process (200).

[0051] The client process (360) sets model enhancement and performs tasks required for model enhancement. The client process (360) can be set to automatically or manually perform model enhancement periodically (automatic setting) or to perform model enhancement using labeled learning data acquired up to the point at which model enhancement is selected (manual setting). Meanwhile, the client process (360) can provide feedback on image data and metadata stored in the detection result storage unit (250) of the image analysis process (200), review labeled learning data by the model enhancement process (300), and view analysis results provided by the monitoring process (370).

[0052] The monitoring process (370) tracks the performance and efficiency of detection models for each channel, and triggers model enhancement of the detection model if improvements to the existing detection model are required. The monitoring process (370) includes technical performance monitoring (371), model performance monitoring (372), and drift detection (373).

[0053] Technical performance monitoring (371) monitors system performance such as system usage and logs of the image analysis process (200), and model performance monitoring (372) monitors the detection model such as FPS (Frames per second) or GPU occupancy of the detection model.

[0054] Drift detection (373) detects changes in the performance of a detection model over time and initiates model advancement. Changes in the performance of a detection model over time can be detected by capturing new data patterns (i.e., drift) from feedback collected by the learning database (310). Changes in the performance of a detection model can occur for various reasons, such as weather or camera conditions. Drift includes concept drift and data drift. Concept drift refers to differences in the characteristics of learned data and predicted data. For example, when only people are detected, only people wearing hats need to be classified and detected, or when only people are detected, vehicles also need to be detected. Data drift refers to differences in the dataset required for detection depending on the weather or environment. For example, since the input data is different depending on various environments and weather, such as snow, rain, sunrise, and sunset, the required data also differs.

[0055] Referring to FIG. 4, in S10, the client process (360) provides feedback on image data and metadata stored in the detection result storage unit (250) of the image analysis process (200). The feedback includes normal detection and error detection, and may additionally or optionally include non-detection.

[0056] In S11, the learning database (310) collects raw learning data stored in the detection result storage unit (250) of the image analysis process (200).

[0057] In S12, drift detection (373) statistically analyzes feedback to detect changes in the detection model's performance over time. The analysis results can also be displayed visually. If changes are detected in real time, model enhancement is initiated. If the number of feedbacks does not meet a threshold, the user may be prompted to initiate the enhancement.

[0058] Figures 5 and 6 are diagrams illustrating exemplary scenarios for deploying advanced detection models.

[0059] When a single image analysis process (200) analyzes image data provided through multiple channels, different detection models may be used for each channel. Detection result data is stored for each channel, and model enhancement may be initiated only for detection models requiring enhancement. Meanwhile, if an error is detected only in the first channel, but there is a possibility that an error may also occur in the second channel in which no error was detected, the detection model corresponding to the second channel may be enhanced using the detection result data from the first channel.

[0060] Referring to FIGS. 5 and 6 together, (a) represents image analysis (200a) using the same detection model for all channels, and (b) represents image analysis (200c) using different detection models for each channel. The numbers in parentheses in the figures represent importance levels.

[0061] In (a), the image analysis (200a) uses an integrated detection model that detects all of falldown, fire, and loitering (Person), and thus analyzes the image data provided through channels CH1 to CH3, and the image analysis (200b) uses an integrated detection model that detects intrusion (Person), and thus analyzes the image data provided through channel CH4. In a multi-channel environment, the same detection model is applied to all channels, so model advancement also targets the integrated detection model. However, the integrated detection model takes a considerable amount of time for model advancement, and it is difficult to obtain a high detection rate because it does not reflect channel-specific characteristics (e.g., background).

[0062] In (b), image analysis (200c, 200d) uses a detection model trained for each channel, enabling detection model advancement for each channel. In particular, relatively rapid model advancement is possible compared to an integrated detection model.

[0063] The foregoing description of the present invention is for illustrative purposes only, and those skilled in the art will readily appreciate that the present invention can be readily modified into other specific forms without altering the technical spirit or essential characteristics of the present invention. Therefore, it should be understood that the embodiments described above are exemplary in all respects and not restrictive. In particular, the features of the present invention described with reference to the drawings are not limited to the structures depicted in the specific drawings, and may be implemented independently or in combination with other features.

[0064] The scope of the present invention is indicated by the claims described below rather than the detailed description above, and all changes or modifications derived from the meaning and scope of the claims and their equivalent concepts should be interpreted as being included in the scope of the present invention.

Claims

1. In an intelligent image analysis system based on automatic learning of situational data, The above image analysis system includes multiple processes that are independently executed on a computing device comprising a central processing unit and memory. The above multiple processes are: An image analysis process for storing, by channel, image data in which one or both of objects and events are detected using detection models provided for each channel, metadata generated in association with one or both of the detected objects and events, and feedback on the image data and the metadata, wherein the feedback includes normal detection and error detection; and An image analysis system comprising a model enhancement process for collecting the image data, the metadata and the feedback, and detecting drift by monitoring the collected feedback, thereby generating an advanced detection model.

2. In claim 1, the model enhancement process comprises: A monitoring process for monitoring feedback from the above image analysis by channel and initiating model advancement of the detection model of the channel where the above drift is detected; A labeling process for generating labeled training data including detected images in which one or both of objects and events are detected from the image data and marked with bounding boxes, and metadata associated with the detected images; A detection model generation process that trains a base model with the above-mentioned labeled learning data to generate an advanced detection model; A model validation process for verifying the performance of the above-mentioned advanced detection model; and An image analysis system comprising a model deployment process for deploying a validated detection model to the image analysis.

3. An image analysis system according to claim 2, wherein the labeled learning data includes first labeled learning data used for learning a detection model of a channel in which the drift is detected and unused second labeled learning data.

4. In claim 3, the model verification process is an image analysis system that commonly applies a verification data set selected from the second labeled learning data to the detection model of the channel in which the drift is detected and the advanced detection model.

5. In claim 1, the detection model generation process trains two or more base models having different computing power requirements using the same labeled learning data, and the detection model requiring relatively high computing power is loaded into the labeling process, and the detection model requiring relatively low computing power is loaded into the image analysis process. An image analysis system.

6. An image analysis system according to claim 1, further comprising a client process that provides feedback on the image data and the metadata stored in the image analysis process.

7. An image analysis system according to claim 1, wherein the image analysis process uses detection models learned for each of a plurality of channels, and only the detection model used for the channel in which the drift is detected is advanced.

Images