CN114157829A - Model training optimization method and device, computer equipment and storage medium - Google Patents
Model training optimization method and device, computer equipment and storage medium Download PDFInfo
- Publication number
- CN114157829A CN114157829A CN202010932795.5A CN202010932795A CN114157829A CN 114157829 A CN114157829 A CN 114157829A CN 202010932795 A CN202010932795 A CN 202010932795A CN 114157829 A CN114157829 A CN 114157829A
- Authority
- CN
- China
- Prior art keywords
- model
- image
- violation
- key frame
- target
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000012549 training Methods 0.000 title claims abstract description 129
- 238000005457 optimization Methods 0.000 title claims abstract description 92
- 238000000034 method Methods 0.000 title claims abstract description 76
- 238000003860 storage Methods 0.000 title claims abstract description 21
- 238000001514 detection method Methods 0.000 claims abstract description 131
- 238000012544 monitoring process Methods 0.000 claims abstract description 62
- 238000012545 processing Methods 0.000 claims description 17
- 238000004590 computer program Methods 0.000 claims description 10
- 230000006399 behavior Effects 0.000 claims description 9
- 238000012216 screening Methods 0.000 claims description 8
- 238000010191 image analysis Methods 0.000 claims description 4
- 238000004458 analytical method Methods 0.000 abstract description 29
- 230000001976 improved effect Effects 0.000 description 22
- 238000010801 machine learning Methods 0.000 description 18
- 238000002372 labelling Methods 0.000 description 12
- 238000005516 engineering process Methods 0.000 description 11
- 238000004519 manufacturing process Methods 0.000 description 10
- 238000013135 deep learning Methods 0.000 description 9
- 238000010586 diagram Methods 0.000 description 6
- 238000000605 extraction Methods 0.000 description 6
- 238000013473 artificial intelligence Methods 0.000 description 5
- 238000013527 convolutional neural network Methods 0.000 description 4
- 238000013480 data collection Methods 0.000 description 4
- 238000013500 data storage Methods 0.000 description 4
- 241000282414 Homo sapiens Species 0.000 description 3
- 230000005540 biological transmission Effects 0.000 description 3
- 238000003062 neural network model Methods 0.000 description 3
- 230000008093 supporting effect Effects 0.000 description 3
- 238000013528 artificial neural network Methods 0.000 description 2
- 230000006835 compression Effects 0.000 description 2
- 230000006870 function Effects 0.000 description 2
- 238000012015 optical character recognition Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000012552 review Methods 0.000 description 2
- 230000000007 visual effect Effects 0.000 description 2
- 241001465754 Metazoa Species 0.000 description 1
- 230000003044 adaptive effect Effects 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000003190 augmentative effect Effects 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 230000001939 inductive effect Effects 0.000 description 1
- 230000010365 information processing Effects 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 230000000306 recurrent effect Effects 0.000 description 1
- 230000002787 reinforcement Effects 0.000 description 1
- 230000001360 synchronised effect Effects 0.000 description 1
- 238000013526 transfer learning Methods 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Images
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N7/00—Television systems
- H04N7/18—Closed-circuit television [CCTV] systems, i.e. systems in which the video signal is not broadcast
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/08—Logistics, e.g. warehousing, loading or distribution; Inventory or stock management
- G06Q10/083—Shipping
Landscapes
- Business, Economics & Management (AREA)
- Engineering & Computer Science (AREA)
- Economics (AREA)
- Marketing (AREA)
- Quality & Reliability (AREA)
- Signal Processing (AREA)
- Entrepreneurship & Innovation (AREA)
- Human Resources & Organizations (AREA)
- Multimedia (AREA)
- Operations Research (AREA)
- Development Economics (AREA)
- Strategic Management (AREA)
- Tourism & Hospitality (AREA)
- Physics & Mathematics (AREA)
- General Business, Economics & Management (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
The application provides a model training optimization method, a device, computer equipment and a storage medium, based on the method, a server receives a monitoring video stream uploaded by at least one front-end device, then key frame images of a target monitoring video in the monitoring video stream are obtained, the key frame images are input into a trained violation behavior detection model, detection results output by the violation behavior detection model are obtained, and finally an image acquisition strategy of the key frame images is optimized based on the detection results, so that the violation behavior detection model is optimally trained through the optimized image acquisition strategy. The method can improve the efficiency of model analysis.
Description
Technical Field
The application relates to the field of artificial intelligence, in particular to a model training optimization method and device for a logistics monitoring system, computer equipment and a storage medium.
Background
With the rapid development of artificial intelligence, the neural network model based on machine learning can identify various information, realize data processing under different application scenes, particularly serve as deep learning of a machine learning subset, and be widely applied to various scenes. For example, image analysis, risk early warning and the like relate to application scenes of computer vision technology, namely, classification and identification of a large amount of data can be realized based on a deep learning technology, so that related classification prediction results can be quickly and accurately obtained, and the function realization of the application scene where the data is located is accelerated.
However, due to the data-driven characteristic of deep learning, a large amount of time is required for data acquisition, manual labeling and the like in analysis work involving computer vision, and the model analysis efficiency is low due to the complicated processing process.
Disclosure of Invention
Based on this, the application provides a model training optimization method, a device, computer equipment and a storage medium for a logistics monitoring system, the logistics monitoring system comprises front-end equipment and a server, the front-end equipment is connected with the server in a wireless connection mode, and the model training optimization method is applied to the server and used for improving machine learning accuracy and further improving model analysis efficiency.
In a first aspect, the present application provides a model training optimization method, including:
receiving a monitoring video stream uploaded by at least one piece of front-end equipment;
acquiring a key frame image of a target surveillance video in the surveillance video stream;
inputting the key frame image into a trained violation detection model to obtain a detection result output by the violation detection model;
and optimizing an image acquisition strategy of the key frame image based on the detection result, and performing optimization training on the violation detection model through the optimized image acquisition strategy.
In a second aspect, the present application provides a model training optimization apparatus, the apparatus comprising:
the video receiving module is used for receiving the monitoring video stream uploaded by at least one piece of front-end equipment;
the image acquisition module is used for acquiring a key frame image of a target monitoring video in the monitoring video stream;
the image analysis module is used for inputting the key frame image into a trained violation detection model to obtain a detection result output by the violation detection model;
and the strategy optimization module is used for optimizing an image acquisition strategy of the key frame image based on the detection result so as to perform optimization training on the violation behavior detection model through the optimized image acquisition strategy.
In a third aspect, the present application further provides a server, including:
one or more processors;
a memory; and
one or more applications, wherein the one or more applications are stored in the memory and configured to be executed by the processor to implement the model training optimization method.
In a fourth aspect, the present application further provides a computer-readable storage medium, on which a computer program is stored, the computer program being loaded by a processor to perform the steps of the model training optimization method.
According to the model training optimization method, the model training optimization device, the computer equipment and the storage medium, the key frame image of the target monitoring video in the monitoring video stream is input into the violation detection model, so that the detection result output by the model can be used for optimizing an image acquisition strategy, and then the optimized strategy is used for carrying out iterative training on the violation detection model. By adopting the method, closed-loop training optimization of the violation detection model is realized, manual analysis on each video in the monitoring video stream is not needed, and only manual rechecking is needed on the detection result corresponding to the key frame image, so that the machine learning accuracy can be improved, the labor cost can be saved, and the model analysis efficiency can be improved.
Drawings
In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the description of the embodiments are briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present application, and it is obvious for those skilled in the art to obtain other drawings based on these drawings without creative efforts.
FIG. 1 is a schematic view of a scenario of a model training optimization method in an embodiment of the present application;
FIG. 2 is a schematic flow chart of a model training optimization method in an embodiment of the present application;
FIG. 3 is a schematic structural diagram of a model training optimization system in an embodiment of the present application;
FIG. 4 is a schematic flowchart of a method for optimizing model training in an embodiment of the present application;
FIG. 5 is a schematic structural diagram of a model training optimization device in an embodiment of the present application;
fig. 6 is a schematic structural diagram of a computer device in an embodiment of the present application.
Detailed Description
The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application, and it is obvious that the described embodiments are only a part of the embodiments of the present application, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present application.
In this application, the word "for example" is used to mean "serving as an example, instance, or illustration". Any embodiment described herein as "for example" is not necessarily to be construed as preferred or advantageous over other embodiments. The following description is presented to enable any person skilled in the art to make and use the invention. In the following description, details are set forth for the purpose of explanation. It will be apparent to one of ordinary skill in the art that the present invention may be practiced without these specific details. In other instances, well-known structures and processes are not shown in detail to avoid obscuring the description of the invention with unnecessary detail. Thus, the present invention is not intended to be limited to the embodiments shown, but is to be accorded the widest scope consistent with the principles and features disclosed herein.
The embodiment of the application provides a model training optimization method and device for a logistics monitoring system, a computer device and a storage medium, which are respectively described in detail below.
Referring to fig. 1, fig. 1 is a schematic view of a scenario of a model training optimization method for a logistics monitoring system according to an embodiment of the present application, where the system may include a front-end device 100 and a server 200, and the front-end device 100 and the server 200 may be connected and communicated through an internet formed by various gateways, such as a wide area network and a local area network, which are not described herein again. The front-end device 100 includes, but is not limited to, an embedded high-definition camera, an industrial personal computer, a high-definition camera, and the like, and is configured to perform data acquisition on a vehicle and a person passing through the front-end device, where the data acquisition includes, but is not limited to, a license plate number of the vehicle (the number may be a fake plate or a fake plate), a license plate type (a blue-bottom license plate of a private car, a yellow-bottom license plate of a truck, and the like), and illegal behaviors of the person. In addition, the front-end device 100 may include, but is not limited to, a portable terminal such as a mobile phone and a tablet, and a fixed terminal such as a computer, a query machine and an advertisement machine, and is a service port that can be used and operated by a user. The server 200 may be an independent server, or may be a server network or a server cluster composed of servers, which includes but is not limited to a computer, a network host, a single network server, a plurality of network server sets, or a cloud server composed of a plurality of servers. Among them, the Cloud server is constituted by a large number of computers or web servers based on Cloud Computing (Cloud Computing).
Those skilled in the art will appreciate that the application environment shown in fig. 1 is only one application scenario related to the present application, and does not constitute a limitation to the application scenario of the present application, and that other application environments may further include more or less computer devices than those shown in fig. 1, for example, only 1 server 200 is shown in fig. 1, and it is understood that the model training optimization system may further include one or more other servers, for example, the recognition server 201, the training server 202, and the like. The training server 202 is used for services such as model training, and the recognition server 201 is used for receiving a recognition request from the front-end device 100 and returning a recognition result, and the like, which is not limited herein. In addition, as shown in fig. 1, the model training optimization system may further include a memory 300 for storing data, for example, various data of the logistics platform, such as logistics transportation information of a logistics node such as a transit point, specifically, express mail information, delivery vehicle information, logistics node information, and the like.
It should be noted that, the scenario diagram of the logistics monitoring system shown in fig. 1 is only an example, and the logistics monitoring system and the scenario described in the embodiment of the present invention are for more clearly illustrating the technical solution of the embodiment of the present invention, and do not form a limitation on the technical solution provided in the embodiment of the present invention.
It should be noted that the Computer Vision technology (CV) used in the present application is a science for studying how to make a machine "see", and more specifically, it refers to that a camera and a Computer are used to replace human eyes to perform machine Vision such as recognition, tracking and measurement on a target, and further perform graphics processing, so that the Computer processing becomes an image more suitable for human eyes to observe or transmitted to an instrument to detect. As a scientific discipline, computer vision research-related theories and techniques attempt to build artificial intelligence systems that can capture information from images or multidimensional data. The computer vision technology generally includes image processing, image Recognition, image semantic understanding, image retrieval, Optical Character Recognition (OCR), video processing, video semantic understanding, video content/behavior Recognition, three-dimensional object reconstruction, 3D technology, virtual reality, augmented reality, synchronous positioning, map construction, and other technologies, and also includes common biometric technologies such as face Recognition and fingerprint Recognition.
Finally, it should be noted that the Machine Learning technique (ML) applied in the present application is a multi-domain cross discipline, and relates to a multi-domain discipline such as probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory, and the like. The special research on how a computer simulates or realizes the learning behavior of human beings so as to acquire new knowledge or skills and reorganize the existing knowledge structure to continuously improve the performance of the computer. Machine learning is the core of artificial intelligence, is the fundamental approach for computers to have intelligence, and is applied to all fields of artificial intelligence. Machine learning and deep learning generally include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and the like.
By utilizing the computer vision technology and the machine learning technology, the accuracy of video classification can be effectively improved, for example, the violation identification scene related by the application needs to identify whether the violation exists in the image/video, which is a two-classification problem. However, the existing computer vision algorithm engineers need to spend a lot of time on complicated links such as data acquisition and data labeling, the efficiency is not high, and in the process, because the data volume is huge, the transmission, storage and arrangement of the data cost and the labor cost are huge. Therefore, the violation detection model with excellent performance is designed, the model is obtained based on iterative training of the model training optimization strategy, and model training optimization is performed by adopting the model training optimization strategy, so that the accuracy of machine learning can be improved, and the model analysis efficiency can be improved.
Fig. 2 is a schematic flow chart of an embodiment of a model training optimization method for a logistics monitoring system in the embodiment of the present application, and this embodiment mainly exemplifies that the method is applied to the server 200 in fig. 1, and specifically may be an identification server. Referring to fig. 2, the model training optimization method specifically includes steps S201 to S204, and specifically includes the following steps:
s201, receiving at least one monitoring video stream uploaded by the front-end equipment.
The video may be a dynamic image composed of a series of continuous images, and the monitoring video may be a dynamic image recorded with a specific monitoring scene.
Specifically, the surveillance video stream may be a real-time surveillance video stream, the surveillance video stream includes the front-end device 100, for example, a video set captured by a high-definition camera, the surveillance video stream includes real-time monitoring of a vehicle passing through the front-end device 100 and real-time monitoring of a pedestrian passing through the front-end device, meanwhile, in the real-time surveillance video stream, corresponding time and information corresponding to an address of the front-end device may be recorded, so that when an illegal act occurs in the real-time surveillance video stream, corresponding recording may be performed on the time and the address where the illegal act occurs, so that a corresponding record may be followed when the illegal act is subsequently handled.
S202, acquiring a key frame image of the target monitoring video in the monitoring video stream.
The target surveillance video may refer to a currently marked surveillance video to be analyzed. For example, the target surveillance video may be a surveillance segment in a logistics operation scene, the surveillance segment may be a dynamic image containing illegal behaviors of target detection objects, the target detection objects may be the illegal behaviors of vehicles, people, and the like, and the target detection objects are objects marked to be analyzed in the model training process.
The key frame image may be a video image frame obtained by performing frame extraction processing and image recognition on the target monitoring video.
It should be noted that, although the target surveillance video may be a marked video to be analyzed, the marked implementation manner includes, but is not limited to, a manner of triggering a mark based on that the front-end device 100 has recorded, is not broadcasted, is broadcasted during broadcasting or has broadcasted, and the like, and may also be a manner of designating a mark for a certain video according to an actual situation, for example, when the front-end device 100 is a portable terminal or a fixed terminal, the video a that is broadcasted in advance on the front-end device 100 is automatically given a preset label by the front-end device 100, so as to mark the video a for analysis and identification by the server to be identified; for another example, when the front-end device 100 is a video camera, a preset label is automatically assigned to the video B not played on the front-end device 100 when the recording is completed, so as to mark the video B for analysis and identification by the to-be-identified server. In summary, the manner in which the server 200 acquires the video according to the mark information is not limited in this application. However, in order to fully describe the actions and processing objects executed in the steps of the present application, in this embodiment and the following embodiments, the target surveillance video may be considered as a video actively acquired by the server 200 from the front-end device 100.
Specifically, after the server 200 acquires the target monitoring video, the image processing tool may perform frame extraction processing on the target monitoring video according to a certain frequency and a certain number. For example, frame extraction is performed by using a tool OpenCV (OpenCV is a cross-platform computer vision and machine learning software library) or a tool ffmpeg (ffmpeg is a set of open-source computer programs that can be used to record and convert digital audio and video, and can convert the digital audio and video into a stream), and the like, the frame extraction frequency may be once a second or multiple times a second, and the frame extraction number may be one frame or multiple frames at a time, which is not limited in the present application.
More specifically, after obtaining the target surveillance video and performing video image frame extraction processing, the server 200 analyzes each video image frame, such as object information recognition and machine inference confidence analysis, so as to screen out a key frame image that can be used as a basis for subsequent processing from all video image frames. The screening basis of the key frame image is as follows: and detecting each frame of picture in the target monitoring video frame by frame, and judging the picture as a key frame image by the identification server when suspected illegal actions of people or articles in certain frames are detected.
And S203, inputting the key frame image into the trained violation detection model to obtain a detection result output by the violation detection model.
The violation detection model may be at least one of a Deep Neural Network model (DNN), a Convolutional Neural Network model (CNN), and a Recurrent Neural Network model (RNN), and may be set in a training server shown in fig. 1, and after the training server performs model training, the recognition server may call the trained violation detection model to analyze and obtain a detection result of the key frame image.
Specifically, the current violation detection model is a violation detection model (including but not limited to CNN, DNN, RNN, and the like) after basic training, and the trained violation detection model has a capability of performing behavior prediction on the model input data — the key frame image, so that the key frame image is input into the trained violation detection model, the trained violation detection model can correspondingly output an image including a violation by using a detection mode learned by early training, and a plurality of images are collected to form a video clip, which is a detection result.
More specifically, the hardware capable of supporting model training is a server cluster with a GPU (graphics Processing unit), the server cluster can schedule a training task initiated by a user according to a preset training strategy (network hyper-parameters, etc.) and resources (video memory, memory) required by the model, and the state (Loss/Accuracy access, video memory occupation, GPU utilization, etc.) of the training process can be fed back to the Web end of the server 200 in real time through the front-end device 100. In addition, as will be appreciated by those skilled in the art, the hyper-parameters are parameters that are set before the learning is started, and are not parameter data obtained through training, and in general, the hyper-parameters need to be optimized, and a set of optimal hyper-parameters is selected to improve the performance and effect of model learning.
It should be noted that after each model training, a new violation detection model may be deployed at the cloud or the edge, that is, a binary executable file corresponding to the violation detection model is associated with the model attribute, and the binary executable file is packaged, pushed and updated (a program package is pushed to an edge server or a cloud server, and the program package may be understood as a compressed package, and may be run after being decompressed, including the executable file and the model file), and a local server records a version change.
And S204, optimizing an image acquisition strategy of the key frame image based on the detection result, and performing optimization training on the violation detection model through the optimized image acquisition strategy.
Specifically, the key frame image is input to the trained violation detection model, after a detection result output by the violation detection model is obtained, the detection result can be fed back to a terminal used by a user, and the user can feed back whether the detection result inferred by the current model is correct or not through a Web page of the terminal, so that a manual detection result based on the machine detection result is obtained.
More specifically, after the server 200 obtains the machine detection result and the manual detection result, it may determine whether the two results are consistent, and if not, may optimize the preset image acquisition policy of the application using the key frame image, for example, collect the sample type with the wrong identification of the production environment back for reuse in training, and the model training policy, for example, update the hyper-parameters of the model, so that the model obtains a more appropriate and accurate training sample, and learns a more optimal identification method.
For example, the server 12 reports an illegal video segment of the relevant person, the administrator can review the video to determine whether the violation is true, and if the recognition result (machine detection result) is inconsistent with the review result (manual detection result) of the administrator, the data of the video segment is returned to the data collector to optimize the image acquisition policy and optimize the model training policy.
In this embodiment, the key frame image of the target surveillance video in the surveillance video stream is input to the violation detection model, so that the detection result output by the model can be used for optimizing an image acquisition strategy, and then the optimized strategy is used for performing iterative training on the violation detection model. By adopting the method, closed-loop training optimization of the violation detection model is realized, manual analysis on each video in the monitoring video stream is not needed, and only manual rechecking is needed on the detection result corresponding to the key frame image, so that the machine learning accuracy can be improved, the labor cost can be saved, and the model analysis efficiency can be improved.
In an embodiment, mainly illustrated by applying the model training optimization method to the recognition server in fig. 1, step S202 specifically includes the following steps:
s301, acquiring a target monitoring video in the monitoring video stream;
s302, carrying out object information identification on a target object in the target monitoring video to obtain object identification data of the target object, wherein the object information comprises an object identity, an action and a coordinate position;
s303, acquiring the key frame image according to the object identification data.
The target object may be a specific person or object in the video image, such as a specific person, a specific car, a specific license plate, and the like. It is understood that, in other embodiments, the target object may also be an item such as a lane number, a traffic light, and the like, and therefore, the target object is not particularly limited in the embodiment of the present application.
Specifically, after receiving the target monitoring video sent by the edge or the cloud, the server 200 may perform object information identification on the target object included in the target monitoring video, acquire identification data of the target object, such as identity data, motion data, coordinate position data locked in tracking, and the like, and acquire a keyframe image in the target monitoring video based on the object identification data.
For example, the method can detect actions of target objects (such as express delivery personnel and logistics workers) in the target monitoring video, for example, detect actions of the express delivery personnel and the logistics workers contacting packages and sorting the packages, and acquire action data of designated personnel as object identification data. The state detection of the target object (transport means), such as recognition of the loading state of express delivery vehicles and commuter vehicles (estimated loading rate), recognition of license plates, and the like, may also be performed, and the state data of the specified items is acquired as the object recognition data.
In the embodiment, the target monitoring video is subjected to object information identification, and the object identification data in the video image is analyzed and obtained for production analysis, so that the requirements of practical application can be met, the machine learning accuracy is improved, and the model analysis efficiency can be improved.
In an embodiment, the model training optimization method is mainly applied to the recognition server in fig. 1 for example, and the step S303 specifically includes the following steps:
s401, when the object identification data meet a preset violation triggering condition, acquiring a violation video frame image of the target object in the target monitoring video;
s402, obtaining confidence coefficient data of the violation video frame image based on an edge-end algorithm;
and S403, acquiring the key frame image according to the confidence coefficient data.
The Confidence (Confidence) is expressed by the formula, Confidence (a — > B) ═ P (a | B), in the deep learning field, and it reveals the probability that the item set B also appears in the transaction set D in which the item set a appears. For example, if B also appears or has a high probability of appearing when a appears, a and B may be bundled if the confidence is 100%. If the confidence is too low, it is indicated whether the occurrence of A is not related to the occurrence of B. It will be appreciated that confidence is used in the model training, with higher confidence indicating more positive results from the model on the output. For example, the confidence threshold is 0.7, which means that if the calculated confidence value exceeds 0.7, the verification experiment is successful, or the calculated result is valid and usable.
Specifically, based on the understanding, the embodiment of the present application provides that when the object identification data of any one target object meets the violation triggering condition, the violation video frame image of the target object in the target monitoring video can be obtained, and then the violation video frame image is identified and inferred through the preset edge server, so as to obtain confidence data of each violation video frame image, and if the confidence data reaches a certain rule, the confidence data can be collected by the data collector at the edge end, and then transmitted to the data storage cluster at the cloud end for storage.
More specifically, the violation triggering condition may be set according to an actual application requirement, that is, by presetting identification data with specified image interception features, for example, action data with personnel violation behavior features, the server 200 may perform object information identification on a target object in the target monitoring video, analyze and match the obtained object identification data, if the identification data with the image interception features exists in the currently obtained object identification data, extract a violation video frame image for a specific time sequence frame position of the target object in the target monitoring video, analyze and infer the violation video frame image through an edge algorithm to output a confidence level, and obtain a key frame image according to a confidence level rule embodied by the confidence level data.
In the embodiment, the key frame image is obtained through confidence coefficient analysis, so that the machine learning accuracy can be improved, and the model analysis efficiency can be improved.
In an embodiment, the model training optimization method is mainly applied to the recognition server in fig. 1 for example, and the step S403 specifically includes the following steps:
s501, determining whether the confidence coefficient data meet preset image screening conditions;
and S502, when the confidence coefficient data meet a preset image screening condition, acquiring the violation video frame image as the key frame image.
Specifically, based on the description of the above embodiment, the image screening condition may refer to a confidence rule, and when the confidence satisfies a certain rule, for example, an increasing trend of the confidence satisfies a preset trend range, a change rate of the confidence meets a preset change degree, and the like, the violation video frame image that meets the condition may be saved as a key frame image.
In the embodiment, the key frame image is obtained through confidence coefficient analysis, so that the machine learning accuracy can be improved, and the model analysis efficiency can be improved.
In an embodiment, the model training optimization method is mainly applied to the training server in fig. 1 as described above for example, and after step S204, the following steps are specifically included:
s601, acquiring a target sample image in the monitoring video stream based on the optimized image acquisition strategy;
s602, determining an illegal action area in the target sample image, wherein the illegal action area is determined according to the object identification data of the target object in the target sample image;
s603, carrying out image annotation on the illegal action area through a cloud end to obtain an annotated target sample image;
and S604, performing optimization training on the violation detection model according to the labeled target sample image to obtain an optimized violation detection model.
The target sample image is a sample image which is subjected to manual detection and authentication and is wrongly identified by the violation detection model in the early stage.
The illegal action area may be an image area in which a preset illegal action and an illegal state exist in the video image.
Specifically, image annotation refers to adding text feature information reflecting the content of an image to the image in a machine learning manner according to the visual content of the image. The method utilizes a labeled image set or other available information to automatically learn the potential association or mapping relation between a semantic concept space and a visual characteristic space, adds text keywords to an unknown image, and can convert an image information problem into a text information processing problem with relatively mature technology through the processing of an image automatic labeling technology.
More specifically, the server 200 may obtain the labeled target sample image by obtaining a labeling result obtained by labeling the target sample image with a feature type by one or more labeling personnel, where the related features may be operation features required to be detected and identified by the model, such as city traffic geographic features (traffic road conditions, traffic lights, traffic signs, etc.), life natural features (people, animals, trees, etc.), and vehicle features (vehicle body, license plate, wheels, etc.).
Furthermore, the annotating personnel can use an annotating tool at the Web end or the desktop end on the front-end device 100 side to annotate the area where the illegal action occurs in the target sample image, the existing algorithm can be used for generating pre-annotation information in the annotating process, and only the annotation generated by the algorithm needs to be finely adjusted in the subsequent annotation, so that the annotation time can be saved, and the image (video) is transmitted to the webpage (client) from the cloud without occupying the storage space of an annotation user. Meanwhile, in the labeling process, the target sample image can be transmitted to the cloud end, so that data labeling can be completed at the cloud end, and model training related to subsequent steps can be completed, and therefore the operation time and the local storage space of the server 200 can be saved.
For example, an image region in the key frame image where the illegal action is determined to occur is marked.
Furthermore, the marked target sample image is obtained based on the optimized image acquisition strategy, so that the violation detection model is subjected to iterative optimization training through the marked target sample image, the model identification accuracy can be improved, and the model analysis efficiency can be improved.
In the embodiment, cloud marking is carried out on the illegal action area in the target sample image, so that the local data storage space can be saved, the machine learning accuracy is improved, and the model analysis efficiency can be improved.
In an embodiment, mainly exemplified by applying the model training optimization method to the training server in fig. 1, step S604 specifically includes the following steps:
s701, obtaining model attribute information of the optimized violation detection model;
s702, according to the model attribute information, determining a target execution file in a plurality of pre-stored candidate execution files, wherein the candidate execution files are respectively associated with corresponding model attribute information;
s703, acquiring the operating package data of the optimized violation detection model based on the target execution file;
and S704, sending the operating package data to an edge server and/or a cloud server for deployment processing so that the edge server and/or the cloud server can operate the optimized violation detection model.
Specifically, the model attribute may refer to an attribute category of the model, and the model attribute information records category information of the model, for example, the CNN, DNN, and RNN have their attribute information respectively, and currently, deployment on the server needs to be implemented by different binary executable files because the basic frameworks of the models are different, and then a plurality of binary executable files, such as compressed packets, which can be used by the server to load the trained model may be pre-stored in the database of the server 200 as candidate executable files. Different candidate execution files respectively have different model data tags, the server 200 can determine the model data tags of the model after obtaining the model attribute information of the violation detection model, further can determine the target execution file matched with the candidate execution file in the pre-stored candidate execution files through matching, then pushes the target execution file (program compression package) to an edge server or a cloud server, and can run after the program compression package is decompressed, wherein the target execution file comprises an executable file and a model file.
In this embodiment, by obtaining the model attribute information of the violation detection model and performing model deployment by using the execution file preset by the model attribute information, the requirements of practical application can be met, and the model analysis efficiency is improved.
In an embodiment, the model training optimization method is mainly applied to the recognition server in fig. 1 for example, and the step S204 specifically includes the following steps:
s801, acquiring an artificial detection result corresponding to the detection result;
s802, when the manual detection result is not matched with the detection result, optimizing an image acquisition strategy of the key frame image based on the key frame image to obtain an optimized image acquisition strategy;
and S803, updating the training strategy of the violation detection model based on the optimized image acquisition strategy, so as to perform optimization training on the violation detection model.
The manual detection result may be a rechecking detection result of the detection result by the user through the edge or the cloud.
Specifically, on the basis of the above embodiment, the trained model is deployed in the edge server or the cloud server, that is, the trained model is deployed in the production environment, the production environment (production analysis service) reports the violation video clip, and the administrator checks the video to determine whether the violation video clip is true. When the recognition result of the algorithm is inconsistent with the rechecking result of the manager, the data (recognition error sample) of the video segment is returned to a data collector in the model training optimization system, namely, the data is added into a new round of training set so as to optimize the image acquisition strategy of the key frame image, and the optimized image acquisition strategy can optimize the model training strategy and can also continuously optimize the machine learning scheme by combining the optimized model training strategy.
In this embodiment, whether the keyframe image corresponding to the current detection result can be used to optimize the image acquisition policy is determined by analyzing the manual detection result and the model detection result, and then the labeled keyframe image is used to optimize the image acquisition policy under the condition that the two results are inconsistent, so that the machine learning accuracy can be improved, and the model analysis efficiency can also be improved.
In order to fully understand the model training optimization scheme proposed by the present application, the present application further provides an application scenario applying the model training optimization method described above. Specifically, the application of the model training optimization method in the application scenario will be described as follows with reference to fig. 3 to 4:
as shown in fig. 3, the present application further provides a model training optimization system, which can satisfy the hardware execution conditions of the model training optimization method, can automatically search data, efficiently store data, generate pre-labeling, automatically train process scheduling, automatically pack, push and deploy, and automatically adjust and optimize according to production performance, and can reduce data storage cost and save data transmission time on the basis of reducing the workload of computer vision algorithm engineers in collecting data, labeling data, and analyzing data.
As shown in fig. 3, the system architecture diagram of the model training optimization system in one embodiment is shown, where the model training optimization system is composed of an application layer, a multi-platform inference layer, a general deep learning platform, middleware, an operating system layer, and a hardware layer, and functions configured for each layer are actually embodied in six stages, including: the method comprises a data collection stage, a labeling stage, a training stage, a deployment stage, an inference stage and a production analysis stage.
The application layer is a module for supporting each service, and the application layer and the multi-platform reasoning layer can cooperatively execute the relevant operation of the reasoning stage described in the model training optimization method; the universal deep learning platform can execute the related operations of the data collection stage, the data labeling stage, the model training stage, the model deployment stage and the production analysis stage described in the method; the middleware, the operating system and the hardware are software and hardware facilities of the system for supporting the operation of the deep learning platform. Therefore, the optimization of the image acquisition strategy can be realized by the back transmission of the detection result.
In a data collection stage, the data collector is used for presenting a certain rule to confidence obtained by inference in an unfamiliar scene during forward inference of a neural Network during system operation, and for such a scene, a Payload application program of an edge computing framework in a multi-platform inference layer automatically packages data and then transmits the data to a data Storage cluster at a cloud end for Storage, such as Network Attached Storage (NAS) and Object Storage Service (OSS for short); in the marking stage, the stored pictures/videos can be marked through a Web end/desktop end marking tool preset by the application, so that the storage space of a local database of a user is saved; in the training stage, as described in the above embodiment, a server cluster with multiple GPUs is used for scheduling model training, and the state of the training process is fed back to the Web server in real time, so that a new model can be obtained after training is completed; after a new model is obtained, a deployment stage is started, in the deployment stage, a universal deep learning platform in the system is adaptive to various platforms such as edge computing and cloud computing, corresponding binary executable files can be associated through model data, the binary executable files are packaged, pushed and updated, and version changes are recorded; in the reasoning stage, a reasoning service can be provided for a user by setting an interface at an edge end or a cloud end, and the user can feed back a model to detect whether a result is correct or not on a Web page; in the production service stage, the production analysis service can periodically analyze the performance of the system such as accuracy and the like, and the closed loop of the whole process is realized by updating the model training strategy and the data collection strategy.
In the embodiment, the key frame images marked in the target monitoring video are input into the violation detection model, so that the detection result output by the model can be used for optimizing the image acquisition strategy, and then the optimized strategy is used for carrying out iterative training on the violation detection model.
In order to better implement the model training optimization method in the embodiment of the present application, on the basis of the model training optimization method, an embodiment of the present application further provides a model training optimization device, as shown in fig. 5, the model training optimization device includes:
a video receiving module 502, configured to receive a monitoring video stream uploaded by at least one of the front-end devices;
an image obtaining module 504, configured to obtain a key frame image of a target surveillance video in the surveillance video stream;
an image analysis module 506, configured to input the keyframe image into a trained violation detection model, so as to obtain a detection result output by the violation detection model;
a policy optimization module 508, configured to optimize an image acquisition policy of the keyframe image based on the detection result, so as to perform optimization training on the violation detection model through the optimized image acquisition policy.
In some embodiments of the present application, the image obtaining module 502 is further configured to obtain a target surveillance video in the surveillance video stream; carrying out object information identification on a target object in the target monitoring video to obtain object identification data of the target object, wherein the object information comprises an object identity, an action and a coordinate position; and acquiring the key frame image according to the object identification data.
In some embodiments of the present application, the image obtaining module 502 is further configured to obtain an illegal video frame image of the target object in the target surveillance video when the object identification data meets a preset illegal triggering condition; obtaining confidence data of the violation video frame image based on an edge-end algorithm; and acquiring the key frame image according to the confidence coefficient data.
In some embodiments of the present application, the image obtaining module 502 is further configured to determine whether the confidence data meets a preset image screening condition; and when the confidence coefficient data meet a preset image screening condition, acquiring the violation video frame image as the key frame image.
In some embodiments of the present application, the image annotation module 504 is further configured to obtain a target sample image in the surveillance video stream based on the optimized image acquisition policy; determining an illegal action area in the target sample image, wherein the illegal action area is determined according to the object identification data of the target object in the target sample image; carrying out image annotation on the illegal action area through a cloud end to obtain an annotated target sample image; and performing optimization training on the violation detection model according to the labeled target sample image to obtain an optimized violation detection model.
In some embodiments of the present application, the model training optimization apparatus 500 further includes a model deployment module, configured to obtain model attribute information of the optimized violation detection model; determining a target execution file in a plurality of pre-stored candidate execution files according to the model attribute information, wherein the candidate execution files are respectively associated with corresponding model attribute information; acquiring the operating program package data of the optimized violation detection model based on the target execution file; and sending the operating package data to an edge server and/or a cloud server for deployment processing so that the edge server and/or the cloud server can operate the optimized violation detection model.
In some embodiments of the present application, the policy optimization module 508 is further configured to obtain an artificial detection result corresponding to the detection result; when the manual detection result is not matched with the detection result, optimizing an image acquisition strategy of the key frame image based on the key frame image to obtain an optimized image acquisition strategy; and updating the training strategy of the violation detection model based on the optimized image acquisition strategy so as to perform optimization training on the violation detection model.
In the embodiment, the key frame image of the target monitoring video in the monitoring video stream is input to the violation detection model, so that the detection result output by the model can be used for optimizing an image acquisition strategy, and the optimized strategy is further used for performing iterative training on the violation detection model. By adopting the method, closed-loop training optimization of the violation detection model is realized, manual analysis on each video in the monitoring video stream is not needed, and only manual rechecking is needed on the detection result corresponding to the key frame image, so that the machine learning accuracy can be improved, the labor cost can be saved, and the model analysis efficiency can be improved.
In some embodiments of the present application, the model training optimization apparatus may be implemented in the form of a computer program that is executable on a computer device such as that shown in fig. 6. The memory of the computer device may store various program modules constituting the model training optimization apparatus, such as an image acquisition module 502, an image annotation module 504, an image input module 506, and a strategy optimization module 508 shown in fig. 5. The program modules constitute computer programs that cause the processors to perform the steps in the model training optimization methods of the various embodiments of the present application described in the present specification.
For example, the computer device shown in fig. 6 may execute step S201 through the image acquisition module 502 in the model training optimization apparatus shown in fig. 5. The computer device may perform step S202 through the image annotation module 504. The computer device may perform step S203 through the image input module 506. The computer device may perform step S204 by the policy optimization module 508. The computer device includes a processor, a memory, and a network interface connected by a system bus. Wherein the processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of an operating system and computer programs in the non-volatile storage medium. The network interface of the computer device is used for communicating with an external computer device through a network connection. The computer program is executed by a processor to implement a model training optimization method.
Those skilled in the art will appreciate that the architecture shown in fig. 6 is merely a block diagram of some of the structures associated with the disclosed aspects and is not intended to limit the computing devices to which the disclosed aspects apply, as particular computing devices may include more or less components than those shown, or may combine certain components, or have a different arrangement of components.
In some embodiments of the present application, there is provided a computer device comprising one or more processors; a memory; and one or more applications, wherein the one or more applications are stored in the memory and configured to be executed by the processor to perform the steps of the model training optimization method described above. Here, the steps of the model training optimization method may be steps in the model training optimization methods of the above embodiments.
In some embodiments of the present application, a computer-readable storage medium is provided, in which a computer program is stored, which is loaded by a processor, so that the processor performs the steps of the above-mentioned model training optimization method. Here, the steps of the model training optimization method may be steps in the model training optimization methods of the above embodiments.
The model training optimization method provided by the embodiment of the present application is described in detail above, and the principle and the implementation of the present invention are explained in this document by applying specific examples, and the description of the above embodiment is only used to help understanding the method and the core idea of the present invention; meanwhile, for those skilled in the art, according to the idea of the present invention, there may be variations in the specific embodiments and the application scope, and in summary, the content of the present specification should not be construed as a limitation to the present invention.
Claims (10)
1. A model training optimization method for a logistics monitoring system is characterized in that the logistics monitoring system comprises a front-end device and a server, the front-end device is connected with the server in a wireless connection mode, the model training optimization method is applied to the server, and the model training optimization method comprises the following steps:
receiving a monitoring video stream uploaded by at least one piece of front-end equipment;
acquiring a key frame image of a target surveillance video in the surveillance video stream;
inputting the key frame image into a trained violation detection model to obtain a detection result output by the violation detection model;
and optimizing an image acquisition strategy of the key frame image based on the detection result, and performing optimization training on the violation detection model through the optimized image acquisition strategy.
2. The model training optimization method of claim 1, wherein the step of obtaining key frame images of target surveillance videos in the surveillance video stream comprises:
acquiring a target monitoring video in the monitoring video stream;
carrying out object information identification on a target object in the target monitoring video to obtain object identification data of the target object, wherein the object information comprises an object identity, an action and a coordinate position;
and acquiring the key frame image according to the object identification data.
3. The model training optimization method of claim 2, wherein the step of obtaining the keyframe image from the object recognition data comprises:
when the object identification data meet a preset violation triggering condition, acquiring a violation video frame image of the target object in the target monitoring video;
obtaining confidence data of the violation video frame image based on an edge-end algorithm;
and acquiring the key frame image according to the confidence coefficient data.
4. The model training optimization method of claim 3, wherein the step of obtaining the keyframe image from the confidence data comprises:
determining whether the confidence coefficient data meets a preset image screening condition;
and when the confidence coefficient data meet a preset image screening condition, acquiring the violation video frame image as the key frame image.
5. The model training optimization method of any one of claims 1 to 4, wherein after the optimizing the image acquisition strategy of the key frame images based on the detection results, the method further comprises:
acquiring a target sample image in the monitoring video stream based on the optimized image acquisition strategy;
determining an illegal action area in the target sample image, wherein the illegal action area is determined according to the object identification data of the target object in the target sample image;
carrying out image annotation on the illegal action area through a cloud end to obtain an annotated target sample image;
and performing optimization training on the violation detection model according to the labeled target sample image to obtain an optimized violation detection model.
6. The model training optimization method of claim 5, wherein after the performing optimization training on the violation detection model according to the labeled target sample image to obtain an optimized violation detection model, the method further comprises:
obtaining model attribute information of the optimized violation detection model;
determining a target execution file in a plurality of pre-stored candidate execution files according to the model attribute information, wherein the candidate execution files are respectively associated with corresponding model attribute information;
acquiring the operating program package data of the optimized violation detection model based on the target execution file;
and sending the operating package data to an edge server and/or a cloud server for deployment processing so that the edge server and/or the cloud server can operate the optimized violation detection model.
7. The model training optimization method of claim 1, wherein the optimizing the image acquisition strategy of the key frame image based on the detection result to optimally train the violation detection model through the optimized image acquisition strategy comprises:
acquiring an artificial detection result corresponding to the detection result;
when the manual detection result is not matched with the detection result, optimizing an image acquisition strategy of the key frame image based on the key frame image to obtain an optimized image acquisition strategy;
and updating the training strategy of the violation detection model based on the optimized image acquisition strategy so as to perform optimization training on the violation detection model.
8. The utility model provides a model training optimizing device for logistics monitoring system, its characterized in that, logistics monitoring system includes front-end equipment and server, front-end equipment passes through wireless connection's mode and connects the server, model training optimizing device set up in the server, the device includes:
the video receiving module is used for receiving the monitoring video stream uploaded by at least one piece of front-end equipment;
the image acquisition module is used for acquiring a key frame image of a target monitoring video in the monitoring video stream;
the image analysis module is used for inputting the key frame image into a trained violation detection model to obtain a detection result output by the violation detection model;
and the strategy optimization module is used for optimizing an image acquisition strategy of the key frame image based on the detection result so as to perform optimization training on the violation behavior detection model through the optimized image acquisition strategy.
9. A computer device, characterized in that the computer device comprises:
one or more processors;
a memory; and
one or more applications, wherein the one or more applications are stored in the memory and configured to be executed by the processor to implement the model training optimization method of any of claims 1 to 7.
10. A computer-readable storage medium, having stored thereon a computer program which is loaded by a processor to perform the steps of the model training optimization method according to any one of claims 1 to 7.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010932795.5A CN114157829A (en) | 2020-09-08 | 2020-09-08 | Model training optimization method and device, computer equipment and storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010932795.5A CN114157829A (en) | 2020-09-08 | 2020-09-08 | Model training optimization method and device, computer equipment and storage medium |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114157829A true CN114157829A (en) | 2022-03-08 |
Family
ID=80460537
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010932795.5A Pending CN114157829A (en) | 2020-09-08 | 2020-09-08 | Model training optimization method and device, computer equipment and storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114157829A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114821936A (en) * | 2022-03-21 | 2022-07-29 | 慧之安信息技术股份有限公司 | Method and device for detecting illegal criminal behaviors based on edge calculation |
| CN115379300A (en) * | 2022-07-27 | 2022-11-22 | 国能龙源环保有限公司 | Auxiliary method and auxiliary device for normatively installing explosive package based on AI (Artificial Intelligence) recognition algorithm |
| CN116189063A (en) * | 2023-04-24 | 2023-05-30 | 青岛润邦泽业信息技术有限公司 | Key frame optimization method and device for intelligent video monitoring |
| CN118135377A (en) * | 2024-05-08 | 2024-06-04 | 海信集团控股股份有限公司 | Model deployment method, terminal side equipment and storage medium |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108419091A (en) * | 2018-03-02 | 2018-08-17 | 北京未来媒体科技股份有限公司 | A kind of verifying video content method and device based on machine learning |
| CN109582793A (en) * | 2018-11-23 | 2019-04-05 | 深圳前海微众银行股份有限公司 | Model training method, customer service system and data labeling system, readable storage medium storing program for executing |
| CN109785361A (en) * | 2018-12-22 | 2019-05-21 | 国网内蒙古东部电力有限公司 | Substation's foreign body intrusion detection system based on CNN and MOG |
| US20190332931A1 (en) * | 2018-04-25 | 2019-10-31 | Fujitsu Limited | Deep neural network training for application program generation |
| CN111091098A (en) * | 2019-12-20 | 2020-05-01 | 浙江大华技术股份有限公司 | Training method and detection method of detection model and related device |
| CN111325119A (en) * | 2020-02-09 | 2020-06-23 | 华瑞新智科技（北京）有限公司 | Video monitoring method and system for safety production |
| CN111325067A (en) * | 2018-12-14 | 2020-06-23 | 北京金山云网络技术有限公司 | Illegal video identification method and device and electronic equipment |
| CN111582073A (en) * | 2020-04-23 | 2020-08-25 | 浙江大学 | Transformer substation violation identification method based on ResNet101 characteristic pyramid |
-
2020
- 2020-09-08 CN CN202010932795.5A patent/CN114157829A/en active Pending
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108419091A (en) * | 2018-03-02 | 2018-08-17 | 北京未来媒体科技股份有限公司 | A kind of verifying video content method and device based on machine learning |
| US20190332931A1 (en) * | 2018-04-25 | 2019-10-31 | Fujitsu Limited | Deep neural network training for application program generation |
| CN109582793A (en) * | 2018-11-23 | 2019-04-05 | 深圳前海微众银行股份有限公司 | Model training method, customer service system and data labeling system, readable storage medium storing program for executing |
| CN111325067A (en) * | 2018-12-14 | 2020-06-23 | 北京金山云网络技术有限公司 | Illegal video identification method and device and electronic equipment |
| CN109785361A (en) * | 2018-12-22 | 2019-05-21 | 国网内蒙古东部电力有限公司 | Substation's foreign body intrusion detection system based on CNN and MOG |
| CN111091098A (en) * | 2019-12-20 | 2020-05-01 | 浙江大华技术股份有限公司 | Training method and detection method of detection model and related device |
| CN111325119A (en) * | 2020-02-09 | 2020-06-23 | 华瑞新智科技（北京）有限公司 | Video monitoring method and system for safety production |
| CN111582073A (en) * | 2020-04-23 | 2020-08-25 | 浙江大学 | Transformer substation violation identification method based on ResNet101 characteristic pyramid |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114821936A (en) * | 2022-03-21 | 2022-07-29 | 慧之安信息技术股份有限公司 | Method and device for detecting illegal criminal behaviors based on edge calculation |
| CN115379300A (en) * | 2022-07-27 | 2022-11-22 | 国能龙源环保有限公司 | Auxiliary method and auxiliary device for normatively installing explosive package based on AI (Artificial Intelligence) recognition algorithm |
| CN116189063A (en) * | 2023-04-24 | 2023-05-30 | 青岛润邦泽业信息技术有限公司 | Key frame optimization method and device for intelligent video monitoring |
| CN116189063B (en) * | 2023-04-24 | 2023-07-18 | 青岛润邦泽业信息技术有限公司 | Key frame optimization method and device for intelligent video monitoring |
| CN118135377A (en) * | 2024-05-08 | 2024-06-04 | 海信集团控股股份有限公司 | Model deployment method, terminal side equipment and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Chen et al. | An edge traffic flow detection scheme based on deep learning in an intelligent transportation system | |
| CN114157829A (en) | Model training optimization method and device, computer equipment and storage medium | |
| US9852363B1 (en) | Generating labeled images | |
| US9251425B2 (en) | Object retrieval in video data using complementary detectors | |
| US9489582B2 (en) | Video anomaly detection based upon a sparsity model | |
| Chen et al. | YOLOv5‐Based Vehicle Detection Method for High‐Resolution UAV Images | |
| CN112037142B (en) | Image denoising method, device, computer and readable storage medium | |
| CN114550053A (en) | Traffic accident responsibility determination method, device, computer equipment and storage medium | |
| Mallela et al. | Detection of the triple riding and speed violation on two-wheelers using deep learning algorithms | |
| CN112016545A (en) | Image generation method and device containing text | |
| CN114663871A (en) | Image recognition method, training method, device, system and storage medium | |
| CN114550051A (en) | Vehicle loss detection method and device, computer equipment and storage medium | |
| CN115830399A (en) | Classification model training method, apparatus, device, storage medium, and program product | |
| Kaur et al. | A systematic review of object detection from images using deep learning | |
| CN114329004A (en) | Digital fingerprint generation method, digital fingerprint generation device, data push method, data push device and storage medium | |
| WO2021138893A1 (en) | Vehicle license plate recognition method and apparatus, electronic device, and storage medium | |
| CN115131826B (en) | Article detection and identification method, and network model training method and device | |
| Gao et al. | Segmentation-free vehicle license plate recognition using CNN | |
| Wang et al. | YOLO-ERF: lightweight object detector for UAV aerial images | |
| CN114638973A (en) | Target image detection method and image detection model training method | |
| CN112434601B (en) | Vehicle illegal detection method, device, equipment and medium based on driving video | |
| Khurat et al. | An open-source based automatic car detection system using iot | |
| Han et al. | A Two‐Branch Pedestrian Detection Method for Small and Blurred Target | |
| Lopez-Fuentes et al. | Bandwidth limited object recognition in high resolution imagery | |
| Reddy et al. | Deep CNN Model for Condition Monitoring of Road Traffic: An Application Of Computer Vision |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20220308 |
|
| RJ01 | Rejection of invention patent application after publication |