WO2023143498A1 - 火焰检测方法、装置、设备和存储介质 - Google Patents

火焰检测方法、装置、设备和存储介质 Download PDF

Info

Publication number
WO2023143498A1
WO2023143498A1 PCT/CN2023/073529 CN2023073529W WO2023143498A1 WO 2023143498 A1 WO2023143498 A1 WO 2023143498A1 CN 2023073529 W CN2023073529 W CN 2023073529W WO 2023143498 A1 WO2023143498 A1 WO 2023143498A1
Authority
WO
WIPO (PCT)
Prior art keywords
flame
tested
image
video
flame detection
Prior art date
Application number
PCT/CN2023/073529
Other languages
English (en)
French (fr)
Inventor
陈录城
孟海秀
谢清涛
姚星星
贾冬冬
赵书玲
Original Assignee
卡奥斯工业智能研究院(青岛)有限公司
海尔数字科技(青岛)有限公司
海尔卡奥斯物联科技有限公司
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by 卡奥斯工业智能研究院(青岛)有限公司, 海尔数字科技(青岛)有限公司, 海尔卡奥斯物联科技有限公司 filed Critical 卡奥斯工业智能研究院(青岛)有限公司
Publication of WO2023143498A1 publication Critical patent/WO2023143498A1/zh

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/22—Matching criteria, e.g. proximity measures
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent

Definitions

  • Embodiments of the present application relate to target detection technologies, for example, to a flame detection method, device, equipment, and storage medium.
  • flame safety detection can be performed on surveillance video based on target detection algorithms.
  • the present application provides a flame detection method, device, equipment and storage medium, so as to realize relatively accurate flame detection in a relatively short period of time based on surveillance video.
  • the embodiment of the present application provides a flame detection method, including:
  • a flame detection result is determined according to the category information.
  • the embodiment of the present application also provides a flame detection device, including:
  • the execution module is configured to input the video to be tested into the pre-trained flame detection model, so that the flame detection model determines the image to be tested according to the video to be tested, and classifies the image to be tested to obtain an output result is the category information of the image to be tested;
  • the determination module is configured to determine the flame detection result according to the category information.
  • the embodiment of the present application also provides a computer device, the device comprising:
  • the storage device is configured to store a program, and when the program is executed by the processor, the processor implements the flame detection method according to any one of the first aspect.
  • the embodiment of the present application also provides a storage medium containing computer-executable instructions, and the computer-executable instructions are used to perform the flame detection as described in any one of the first aspect when executed by a computer processor method.
  • the present application provides a computer program product, the computer program product includes computer instructions, and when the computer instructions are run on a computer, the computer executes the flame detection method as provided in the first aspect.
  • all or part of the above computer instructions may be stored on a computer-readable storage medium.
  • the computer-readable storage medium may be packaged together with the processor of the flame detection device, or may be packaged separately with the processor of the flame detection device.
  • Fig. 1 is a flow chart of a flame detection method provided in Embodiment 1 of the present application;
  • FIG. 2 is a flow chart of a flame detection method provided in Embodiment 2 of the present application.
  • FIG. 3 is a schematic diagram of a flame detection model in a flame detection method provided in Embodiment 2 of the present application;
  • FIG. 4 is a schematic diagram of training a flame detection model in a flame detection method provided in Embodiment 2 of the present application;
  • FIG. 5 is a schematic structural diagram of a flame detection device provided in Embodiment 3 of the present application.
  • FIG. 6 is a schematic structural diagram of a computer device provided in Embodiment 4 of the present application.
  • first and second in the specification and drawings of the present application are used to distinguish different objects, or to distinguish different processes for the same object, rather than to describe a specific sequence of objects.
  • the deep learning model includes general target detection and specific target detection.
  • the target detector can use the deep learning network as the backbone and detection network to extract features from the input image or video and then perform classification and positioning to achieve target detection.
  • Mainstream target detectors can usually be divided into two categories, one is two-stage detectors, such as Faster R-CNN; the other is one-stage detectors, such as YOLO and SSD.
  • the one-stage detector has high inference speed, and the two-stage detector has high localization and object recognition accuracy.
  • cameras can be deployed in potential fire-prone locations, and flame safety detection can be carried out in a 24/7 intelligent monitoring method, so as to eliminate the fire in the bud.
  • the present application proposes a flame detection method to achieve high-precision and high-real-time flame detection.
  • FIG. 1 is a flow chart of a flame detection method provided in Embodiment 1 of the present application. This embodiment is applicable to the situation of realizing fast and accurate flame detection based on surveillance video. This method can be executed by a flame detection device, as shown in FIG. 1, including the following steps:
  • Step 110 input the video to be tested into the pre-trained flame detection model, so that the flame detection model determines the image to be tested according to the video to be tested, and classifies the image to be tested, and the output result obtained is Describe the category information of the image to be tested.
  • the flame detection model can include a screening module and a detection module; the screening module can filter the video frames contained in the video to be tested based on the video frame screening method of the related art, so as to determine the image to be tested according to the video to be tested; The module can classify the image to be tested based on the deep learning target detection algorithm to determine the category information of the image to be tested.
  • a surveillance video may be obtained from a camera at a facility where fire is prone to occur, and the surveillance video is determined as the video to be tested, and then the video to be tested is input into the flame detection model as input information.
  • the flame detection model can first filter multiple video frames contained in the video to be tested based on the screening module, and determine the video frame that may contain flames as the image to be tested; then input the image to be tested into the detection module, and the detection module can use the image to be tested Classification is carried out to realize the classification and location of the image to be tested.
  • the output information of the flame detection model may be category information of the image to be tested, and the flame detection model realizes flame detection of the video to be tested.
  • the flame detection model first screens multiple video frames of the video to be tested based on the screening module, and then inputs the screened image to be tested into the detection module to realize the filtering of video frames and screen all video frames that may have flames , reducing the amount of data input to the detection module and improving the detection efficiency of flame detection.
  • the detection module can perform flame detection based on the input image to be tested, and obtain a flame detection result corresponding to the video to be tested.
  • the pre-trained flame detection model can combine the video frame screening method of the related technology with the target detection method of deep learning, and use the video frame filtering method of the related technology before performing target detection on the video frame Perform video frame filtering on the video frames contained in the video to be tested, screen all video frames that may contain flames, and determine the video frames that may contain flames as images to be tested, reduce the amount of data for flame detection, and improve the detection of flame detection efficiency.
  • video frame filtering on the video frames contained in the video to be tested, screen all video frames that may contain flames, and determine the video frames that may contain flames as images to be tested, reduce the amount of data for flame detection, and improve the detection of flame detection efficiency.
  • Step 120 determine the flame detection result according to the category information.
  • the category information may include: the image to be tested contains flames and the image to be tested does not contain flames.
  • the category information is that the image to be tested contains a flame
  • it can be determined that the flame detection result is that there is a flame at the position of the image acquisition device that has acquired the video to be tested; if the category information is that the image to be tested is If the image does not contain a flame, it is determined that the flame detection result is that there is no flame at the position of the image acquisition device that has acquired the video to be tested.
  • the flame detection result of the image to be tested can be determined according to the category information of the image to be tested determined by the flame detection model, and then the flame detection result of the video to be tested to which the image to be tested belongs can be determined.
  • the flame detection result of the position where the device is located is acquired by the image of the measurement video, so as to realize the monitoring of the position prone to fire and reduce the probability of fire occurrence.
  • a flame detection method provided in Embodiment 1 of the present application includes: inputting a video to be tested into a pre-trained flame detection model, so that the flame detection model determines an image to be tested according to the video to be tested, and The image to be tested is classified, and the output result obtained is the category information of the image to be tested; the flame detection result is determined according to the category information.
  • the pre-trained flame detection model can determine the image to be tested according to the video to be tested, reduce the amount of data for target detection, and can also classify the image to be tested based on target detection, and determine the flame detection result according to the classification result, which improves the For the perception ability of flame, the performance of flame detection is also improved, and the real-time and reliability of flame detection are improved at the same time.
  • FIG. 2 is a flow chart of a flame detection method provided in Embodiment 2 of the present application. This embodiment is modified on the basis of the above embodiments. As shown in Figure 2, in this embodiment, the method may also include:
  • Step 210 pre-training the preset target detection model based on the public image set to obtain a pre-training model.
  • the public image set may be an ImageNet dataset, which is a computer vision dataset that includes multiple images and image categories corresponding to the multiple images.
  • a pre-trained target detection model can be pre-trained on a public image set to obtain a pre-trained model.
  • a pre-trained target detection model can be pre-trained on an ImageNet dataset to obtain a pre-trained model.
  • the pre-trained model obtained by pre-training the preset target detection model on the ImageNet dataset has low accuracy for flame detection, and the pre-trained model needs to be migrated. Image set for training again.
  • a first parameter may be determined, and the first parameter may be used to determine a flame detection model.
  • the first parameter When migrating the pre-trained model, the first parameter can be frozen, and other parameters of the pre-trained model can be retrained.
  • the frozen first parameters can be used to determine the flame detection model.
  • the selection of the first parameter may be determined according to actual requirements.
  • the preset target detection model can be pre-trained on the ImageNet dataset to obtain the pre-training model, which solves the flame detection model caused by the lack of images in the training set Difficult to train accurately.
  • Step 220 perform feature matching on multiple flame images included in the flame image set, and determine real labels of the multiple flame images.
  • the image features of multiple flame images contained in the flame image set can be extracted through the feature extraction network; secondly, the anchor frame of the flame image corresponding to the image feature can be quickly generated based on the free anchor frame (Anchor-free) mechanism, And the anchor frame is used for model training, while reducing the number of parameters of the model; furthermore, the feature matching of the anchor frame and the label can be performed based on the sample efficient matching scheme (such as the SimOTA scheme).
  • the label here can be a real label, ie flame.
  • the Anchor-free mechanism significantly reduces the number of design parameters that require heuristic adjustment and many tricks involved (for example, anchor clusters, grid sensitivity) compared to the anchor box-based model, making the flame detection model
  • the training and decoding phases become quite simple.
  • the SimOTA solution automatically sets different numbers of positive samples for flame targets, improving the efficiency of flame detection. Adaptively find the real flame frame that the current feature point should fit through the cost matrix. The higher the coincidence degree, the more fitting is required, the more accurate the classification is, the more fitting is required, and the more fitting is required within a certain radius.
  • Fig. 3 is a schematic diagram of a flame detection model in a flame detection method provided in Embodiment 2 of the present application.
  • the flame detection model includes a detection module, and the detection module includes: Extraction unit, free anchor frame (Anchor-free) unit and sample efficient matching (SimOTA) unit, correspondingly, step 220 may include:
  • Fig. 4 is a schematic diagram of training a flame detection model in a flame detection method provided in Embodiment 2 of the present application.
  • the feature information of a plurality of flame images contained in the flame image set can be determined by a feature extraction unit; Secondly, the anchor frame of the flame image corresponding to the feature information can be quickly generated through the Anchor-free unit; furthermore, the feature matching of the anchor frame can be performed using the SimOTA unit to determine the real label of the flame image corresponding to the anchor frame. Real labels here can be flamed.
  • the YOLOX network (a high-performance detector) needs to perform cluster analysis before training to determine a set of optimal anchor points (anchors).
  • Those aggregated anchors are domain-specific and less general, and increase the complexity of the detection head, as well as the number of predictions per image.
  • AI edge artificial intelligence
  • Anchor-free units can be used to determine anchor boxes.
  • the way to switch the YOLO network to Anchor-free unit is very simple, which can reduce the prediction of each position from 3 to 1, and make them directly predict four values, that is, the two values in the upper left corner of the grid.
  • the application of the anchor-free unit reduces the parameters and computational power of the flame detection model, but achieves better performance.
  • the SimOTA unit can first compute the pairwise matching of anchor boxes and ground truth, denoted by the loss for each prediction-groundtruth pair.
  • the loss between groundtruth and prediction is Among them, ⁇ is the balance coefficient, is the category loss, is the regression loss.
  • the set prediction with the lowest loss in the fixed central region can be selected as its positive sample.
  • the corresponding grids of these positive predictions are assigned positive values, while the remaining grids are negative values.
  • the determination of the number of positive samples can be determined according to the ground truth.
  • the SimOTA unit not only reduces the training time of the flame detection model, but also avoids additional hyperparameter solving.
  • the SimOTA unit increases the average precision (Average Precision, AP) of the flame detection model from 45.0% to 47.3%, improving the accuracy of the flame detection model.
  • the Anchor-free unit in the process of training the flame detection model, is used to solve the problem of a large number of redundant parameters based on the anchor frame mechanism, and the efficiency of model training is improved; the SimOTA unit is used to alleviate the problem of extreme aspect ratio.
  • the problem of poor detection effect of objects and the problem of unbalanced positive samples of extreme-sized objects can be solved, while the problem of poor detection effect of rotating objects can be alleviated, and the efficiency of model training can be improved.
  • Step 230 using the multiple flame images included in the flame image set and the real labels of the multiple flame images as training data to perform network training on the pre-training model, and calculate a loss function.
  • the pre-training model can perform target regression and recognition to determine the training category of the flame images, and compare the The training category and the real category determined by the real label, and calculate the loss function based on the comparison result.
  • the loss function can be selected according to actual needs.
  • Step 240 Perform network optimization based on the backpropagation algorithm until the loss function converges to obtain the flame detection model.
  • network optimization may be performed based on the backpropagation algorithm until the loss function converges.
  • the loss function converges, indicating that the training of the pre-training model is completed, and then the pre-training model whose loss function converges can be determined as the flame detection model.
  • the network is optimized based on the backpropagation algorithm, and when the loss function converges, the second parameter is determined, and the second parameter can also be used to determine the flame detection model. Furthermore, the flame detection model can be determined according to the first parameter determined when the pre-training model is determined and the second parameter determined when the loss function converges.
  • the pre-trained target detection model is trained on the ImageNet dataset to obtain a pre-trained model, and then transfer learning technology is introduced to migrate the pre-trained model, and some parameters are frozen at the same time .
  • the SimOTA unit can be used to determine the real labels of multiple flame images contained in the flame image set, so as to complete the training of the flame detection model in a shorter time.
  • most of the existing target detection algorithms are based on the anchor frame mechanism. Such a design requires a large number of parameters, and the parameters that need to be manually set include: scale and aspect ratio.
  • the scale and aspect ratio in the parameters are more difficult to design.
  • Using the Anchor-free unit to determine the anchor frame can largely avoid the above problems, provide a larger and more flexible solution space, and get rid of the amount of calculation caused by using the anchor frame, improving the accuracy of model training.
  • Step 250 input the video to be tested into the pre-trained flame detection model, so that the flame detection model determines the image to be tested according to the video to be tested, and classifies the image to be tested, and the output result obtained is Describe the category information of the image to be tested.
  • the flame detection model further includes a screening module, and the detection module further includes a main body unit.
  • the flame detection model determines the image to be tested according to the video to be tested, And classifying the image to be tested, the output result obtained is the category information of the image to be tested, including:
  • the subject unit can be the subject detection network of YOLOX.
  • the screening module may screen a large number of video frames included in the video to be tested based on a faster screening method, and determine images that may contain flames as images to be tested.
  • the feature information of the image to be tested can be determined based on the feature extraction unit included in the detection module, and the anchor frame of the image to be tested corresponding to the feature information can be determined based on the Anchor-free unit included in the detection module.
  • the anchor frame can be input into the subject detection network, and the subject detection network can perform target regression and recognition to determine the category information of the image to be tested.
  • the screening module screens the video frames contained in the video to be tested based on an edge detection algorithm to obtain the image to be tested, including:
  • the edge detection algorithm includes a dynamic threshold segmentation algorithm, a mean filter algorithm, and a threshold segmentation algorithm, and accordingly, the screening module is used to analyze the plurality of images contained in the video to be tested based on the edge detection algorithm.
  • Image classification is performed on video frames to obtain classification results, including:
  • the mean filter algorithm and the threshold segmentation algorithm respectively determine the gray value of the video frame; the gray value of the video frame determined based on the dynamic threshold segmentation algorithm value, the grayscale value of the video frame determined based on the mean filtering algorithm, and the base performing weighted summation on the grayscale value of the video determined by the threshold segmentation calculation to obtain the target grayscale value of the video frame; determining the classification of the video frame according to the target grayscale value result.
  • the flame detection model can combine the video frame screening method of the related technology with the target detection method of deep learning.
  • the included video frames are filtered.
  • image classification can be performed on the video frames based on the edge detection algorithm, and the classification result is to screen all the video frames that may have flames, and then determine the video frames that may have flames as the image to be tested. , reduce the amount of flame detection data, and improve the detection efficiency of flame detection.
  • Step 260 determine the flame detection result according to the category information.
  • step 260 may include:
  • the category information is that the image to be tested contains a flame, it is determined that the flame detection result is that there is a flame at the position of the image acquisition device that has acquired the video to be tested; if the category information is that the image to be tested is If the image does not contain a flame, it is determined that the flame detection result is that there is no flame at the position of the image acquisition device that has acquired the video to be tested.
  • the flame detection result of the image to be tested can be determined according to the category information of the image to be tested determined by the flame detection model, and then the flame detection result of the video to be tested to which the image to be tested belongs can be determined.
  • the flame detection result of the position where the device is located is acquired by the image of the measurement video, so as to realize the monitoring of the position prone to fire and reduce the probability of fire occurrence.
  • a flame detection method provided in Embodiment 2 of the present application includes: pre-training a preset target detection model based on a public image set to obtain a pre-trained model; performing feature matching on multiple flame images contained in the flame image set, and determining the multiple the real labels of the flame images; using the multiple flame images contained in the flame image set and the real labels of the multiple flame images as training data to carry out network training on the pre-training model, and calculate the loss function; Perform network optimization based on the backpropagation algorithm until the loss function converges to obtain the flame detection model; input the video to be tested into the pre-trained flame detection model, so that the flame detection model is determined according to the video to be tested The image to be tested is classified, and the output result obtained is the category information of the image to be tested; and the flame detection result is determined according to the category information.
  • the pre-trained target detection model is trained on the ImageNet data set to obtain a pre-trained model, and then transfer learning technology is introduced to migrate the pre-trained model, and the pre-trained model is retrained on the flame image set to obtain a flame detection model.
  • the SimOTA unit can be used to determine the real labels of multiple flame images contained in the flame image set, so as to complete the training of the flame detection model in a shorter time.
  • the Anchor-free unit to determine the anchor frame can provide a larger and more flexible solution space, and get rid of the amount of calculation caused by using the anchor frame, and improve the accuracy of model training.
  • the trained flame detection model can combine the video frame screening method of related technologies with the target detection method of deep learning.
  • the video frames are filtered, and the video frames that may have flames are determined as the images to be tested, so as to reduce the amount of flame detection data and improve the detection efficiency of flame detection. improve The ability to perceive the flame is improved, the performance of the flame detection is also improved, and the real-time and reliability of the flame detection are improved at the same time.
  • FIG. 5 is a schematic structural diagram of a flame detection device provided in Embodiment 3 of the present application.
  • the device can be applied to realize fast and accurate flame detection based on surveillance video, and improve the real-time and accuracy of flame detection.
  • the device can be realized by software and/or hardware, and is generally integrated in computer equipment.
  • the device includes:
  • the execution module 510 is configured to input the video to be tested into the pre-trained flame detection model, so that the flame detection model determines the image to be tested according to the video to be tested, and classifies the image to be tested to obtain an output
  • the result is category information of the image to be tested
  • the determination module 520 is configured to determine the flame detection result according to the category information.
  • the flame detection device inputs the video to be tested into the pre-trained flame detection model, so that the flame detection model determines the image to be tested according to the video to be tested, and classifies the image to be tested , the obtained output result is the category information of the image to be tested; determine the flame detection result according to the category information.
  • the pre-trained flame detection model can determine the image to be tested according to the video to be tested, reduce the amount of data for target detection, and can also classify the image to be tested based on target detection, and determine the flame detection result according to the classification result, which improves the For the perception ability of flame, the performance of flame detection is also improved, and the real-time and reliability of flame detection are improved at the same time.
  • the flame detection model is obtained in the following manner:
  • the pre-training model is obtained;
  • Network optimization is performed based on a backpropagation algorithm until the loss function converges to obtain the flame detection model.
  • the flame detection model includes a detection module, and the detection module includes: a feature extraction unit, a free anchor frame unit and a sample efficient matching unit,
  • the flame detection model also includes a screening module, and the detection module also includes a main unit,
  • the flame detection model determines the image to be tested according to the video to be tested, and classifies the image to be tested, and the output result obtained is the category information of the image to be tested, including:
  • the passing screening module screens the video frames included in the video to be tested based on an edge detection algorithm to obtain the image to be tested, including:
  • the video frame containing flame is determined as the image to be tested by the classification result.
  • the edge detection algorithm includes a dynamic threshold segmentation algorithm, a mean filtering algorithm, and a threshold segmentation algorithm,
  • the mean filter algorithm and the threshold segmentation algorithm respectively determine the gray value of the video frame
  • the grayscale value of the video frame determined based on the dynamic threshold segmentation algorithm For the grayscale value of the video frame determined based on the dynamic threshold segmentation algorithm, the grayscale value of the video frame determined based on the mean filtering algorithm, and the grayscale value determined based on the threshold segmentation algorithm The grayscale value of the video is weighted and summed to obtain the target grayscale value of the video frame;
  • the determination module 520 is configured to determine the classification result of the video frame according to the target gray value in the following manner:
  • the flame detection result is that there is no flame at a location where the image acquisition device that has acquired the video to be tested is located.
  • the flame detection device provided in the embodiment of the present application can execute the flame detection method provided in any embodiment of the present application, and has corresponding functional modules and beneficial effects for executing the method.
  • FIG. 6 is a schematic structural diagram of a computer device provided in Embodiment 4 of the present application.
  • Figure 6 shows a block diagram of an exemplary computer device 6 suitable for implementing embodiments of the present application.
  • the computer device 6 shown in FIG. 6 is only an example.
  • computer device 6 takes the form of a general-purpose computing device.
  • Components of computer device 6 may include at least one processor or processing unit 16 , system memory 28 , bus 18 connecting various system components including system memory 28 and processing unit 16 .
  • Bus 18 represents one or more of several types of bus structures, including a memory bus or memory controller, a peripheral bus, an accelerated graphics port, a processor, or a local bus using any of a variety of bus structures.
  • bus structures include, for example, the Industry Standard Architecture (International Standards Association, ISA) bus, the Micro Channel Architecture (Micro Channel Architecture, MCA) bus, the Enhanced ISA bus, the Video Electronics Standards Association (VESA) ) Local bus and Peripheral Component Interconnect (PCI) bus.
  • Computer device 6 typically includes a variety of computer system readable media. These media can be any available media that can be accessed by the computer device 6 and include both volatile and non-volatile media, removable and non-removable media.
  • System memory 28 may include computer system readable media in the form of volatile memory, such as Random Access Memory (RAM) 30 and/or cache memory 32 .
  • Computer device 6 may include other removable/non-removable, volatile/nonvolatile computer system storage media.
  • storage system 34 may be used to read from and write to non-removable, non-volatile magnetic media (commonly referred to as a "hard drive”).
  • Disk drives for reading and writing to removable non-volatile disks (such as "floppy disks") and for removable non-volatile optical disks (such as Compact Disc-Read Only Memory (CD) -ROM), read-only digital laser disc player (Digital Video Disc-Read Only Memory, DVD-ROM) or other optical media) CD-ROM drive.
  • each drive may be connected to bus 18 via at least one data medium interface.
  • System memory 28 may include at least one program product having a set (eg, at least one) of program modules configured to perform the functions of various embodiments of the present application.
  • a program/utility 40 having a set (at least one) of program modules 42 may be stored, for example, in system memory 28, such program modules 42 including an operating system, at least one application program, other program modules, and program data, in these examples Each or some combination of these may include implementations of network environments.
  • the program modules 42 generally perform the functions and/or methods of the embodiments described herein.
  • Computer device 6 may also communicate with at least one external device 14 (e.g., a keyboard, pointing device, display 24, etc.), and at least one device that enables a user to 6. Any device capable of communicating with one or more other computing devices (eg, network card, modem, etc.). This communication can be performed through an input/output (Input/Output, I/O) interface 22 . And, the computer device 6 can also communicate with at least one network through the network adapter 20 (eg, local area network (Local Area Network, LAN), wide area network (Wide Area Network, WAN) and/or public network, such as the Internet) communication. As shown in FIG. 6 , network adapter 20 communicates with other modules of computer device 6 via bus 18 .
  • network adapter 20 eg, local area network (Local Area Network, LAN), wide area network (Wide Area Network, WAN) and/or public network, such as the Internet
  • computer device 6 may be used in conjunction with computer device 6, including: microcode, device drivers, redundant processing units, external disk drive arrays, Redundant Arrays of Independent Disks (RAID) systems, tape drives And data backup storage system, etc.
  • RAID Redundant Arrays of Independent Disks
  • the processing unit 16 executes various functional applications and page display by running the program stored in the system memory 28, such as implementing the flame detection method provided in the embodiment of the present application, the method includes:
  • a flame detection result is determined according to the category information.
  • processor can also implement the technical solution of the flame detection method provided in any embodiment of the present application.
  • Embodiment 5 of the present application provides a computer-readable storage medium, and a computer program is stored on the computer-readable storage medium.
  • the computer program is executed by a processor, for example, the flame detection method provided by the embodiment of the present invention is implemented.
  • the method includes :
  • a flame detection result is determined according to the category information.
  • the computer storage medium in the embodiments of the present application may use a combination of one or more computer-readable media.
  • the computer readable medium may be a computer readable signal medium or a computer readable storage medium.
  • the computer-readable storage medium may be, for example, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or device, or a combination thereof.
  • Examples of computer readable storage media include: an electrical connection having at least one lead, a portable computer disk, a hard disk, random access memory (RAM), read only memory (ROM), erasable programmable read only memory (such as electronic programmable read-only memory (Electronic Programable Read Only Memory, EPROM) or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or a suitable combination of the above.
  • a computer-readable storage medium may be a tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.
  • a computer readable signal medium may include a data signal carrying computer readable program code in baseband or as part of a carrier wave. This propagated data signal can employ a variety of The form includes electromagnetic signal, optical signal or a suitable combination of the above.
  • a computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in conjunction with an instruction execution system, apparatus, or device. .
  • the program code contained on the computer readable medium can be transmitted by any appropriate medium, including but not limited to: wireless, electric wire, optical cable, radio frequency (Radio Frequency, RF), etc., or a suitable combination of the above.
  • Computer program codes for performing the operations of the present application may be written in at least one programming language or a combination thereof, including object-oriented programming languages such as Java, Smalltalk, C++, and conventional procedural programming A language, such as "C" or a similar programming language.
  • the program code may execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server.
  • the remote computer can be connected to the user computer through any kind of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (such as through the Internet using an Internet service provider). connect).
  • LAN local area network
  • WAN wide area network
  • connect such as AT&T, MCI, Sprint, EarthLink, MSN, GTE, etc.
  • each module or each step of the above-mentioned application can be realized by a general-purpose computing device, and they can be concentrated on a single computing device, or distributed on a network formed by multiple computing devices.
  • they can be implemented with executable program codes of computer devices, so that they can be stored in storage devices and executed by computing devices, or they can be made into individual integrated circuit modules, or a plurality of modules in them Or the steps are fabricated into a single integrated circuit module to realize.
  • the application is not limited to any specific combination of hardware and software.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Evolutionary Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Health & Medical Sciences (AREA)
  • Fire-Detection Mechanisms (AREA)

Abstract

本申请公开了一种火焰检测方法、装置、设备和存储介质，该方法包括：将待测视频输入预先训练好的火焰检测模型，以使所述火焰检测模型根据所述待测视频确定待测图像，并对所述待测图像进行分类，得到的输出结果为所述待测图像的类别信息；根据所述类别信息确定火焰检测结果。

Description

火焰检测方法、装置、设备和存储介质
本公开要求在2022年1月27日提交中国专利局、申请号为202210097095.8的中国专利申请的优先权，该申请的全部内容通过引用结合在本申请中。
技术领域
本申请实施例涉及目标检测技术，例如涉及一种火焰检测方法、装置、设备和存储介质。
背景技术
火灾是给人民大众带来巨大损失的灾难，因此火焰检测是所有的安防系统的重要部分。传统的火焰检测主要基于温度、红外、烟雾等传感器，虽然这些方法成本低廉而且实施简单，但是不能够应用于所有的火灾情况，因为它们的检测范围有限而且不够智能。
近年来，随着监控摄像机的普及以及计算机视觉的发展，人们开始提出一些基于监控视频的火焰检测解决方案，例如可以基于目标检测算法对监控视频进行火焰安全检测。
但是相关技术的基于目标检测算法进行火焰检测的准确性较低，且检测速度也较低，在准确度和实时性上无法满足工业生产需求，难以投入到实际应用中。
发明内容
本申请提供一种火焰检测方法、装置、设备和存储介质，以基于监控视频在较短时间内实现较为准确的火焰检测。
第一方面，本申请实施例提供了一种火焰检测方法，包括：
将待测视频输入预先训练好的火焰检测模型，以使所述火焰检测模型根据所述待测视频确定待测图像，并对所述待测图像进行分类，得到的输出结果为所述待测图像的类别信息；
根据所述类别信息确定火焰检测结果。
第二方面，本申请实施例还提供了一种火焰检测装置，包括：
执行模块，设置为将待测视频输入预先训练好的火焰检测模型，以使所述火焰检测模型根据所述待测视频确定待测图像，并对所述待测图像进行分类，得到的输出结果为所述待测图像的类别信息；
确定模块，设置为根据所述类别信息确定火焰检测结果。
第三方面，本申请实施例还提供了一种计算机设备，所述设备包括：
处理器；
存储装置，设置为存储程序，在所述程序被所述处理器执行时，所述处理器实现如第一方面中任一所述的火焰检测方法。
第四方面，本申请实施例还提供了一种包含计算机可执行指令的存储介质，所述计算机可执行指令在由计算机处理器执行时用于执行如第一方面中任一所述的火焰检测方法。
第五方面，本申请提供一种计算机程序产品，该计算机程序产品包括计算机指令，当计算机指令在计算机上运行时，使得计算机执行如第一方面提供的火焰检测方法。
需要说明的是，上述计算机指令可以全部或者部分存储在计算机可读存储介质上。其中，计算机可读存储介质可以与火焰检测装置的处理器封装在一起的，也可以与火焰检测装置的处理器单独封装。
本申请中第二方面、第三方面、第四方面以及第五方面的描述，可以参考第一方面的描述；并且，第二方面、第三方面、第四方面、以及第五方面的描述的有益效果，可以参考第一方面的有益效果分析。
在本申请中，上述火焰检测装置的名字对设备或功能模块本身不构成限定，在实际实现中，这些设备或功能模块可以以其他名称出现。只要各个设备或功能模块的功能和本申请类似，属于本申请权利要求及其等同技术的范围之内。
本申请的这些方面或其他方面在以下的描述中会更加简明易懂。
附图说明
图1为本申请实施例一提供的一种火焰检测方法的流程图；
图2为本申请实施例二提供的一种火焰检测方法的流程图；
图3为本申请实施例二提供的一种火焰检测方法中火焰检测模型的示意图；
图4为本申请实施例二提供的一种火焰检测方法中火焰检测模型的训练示意图；
图5为本申请实施例三提供的一种火焰检测装置的结构示意图；
图6为本申请实施例四提供的一种计算机设备的结构示意图。
具体实施方式
下面结合附图和实施例对本申请作说明。可以理解的是，此处所描述的实施例仅仅用于解释本申请。另外还需要说明的是，为了便于描述，附图中仅示 出了与本申请相关的部分。
本文中术语“和/或”，仅仅是一种描述关联对象的关联关系，表示可以存在三种关系，例如，A和/或B，可以表示：单独存在A，同时存在A和B，单独存在B这三种情况。
本申请的说明书以及附图中的术语“第一”和“第二”等是用于区别不同的对象，或者用于区别对同一对象的不同处理，而不是用于描述对象的特定顺序。
此外，本申请的描述中所提到的术语“包括”和“具有”以及它们的任何变形，意图在于覆盖不排他的包含。例如包含了一系列步骤或单元的过程、方法、系统、产品或设备没有限定于已列出的步骤或单元，而是可选的还包括其他没有列出的步骤或单元，或可选的还包括对于这些过程、方法、产品或设备固有的其它步骤或单元。
在讨论示例性实施例之前应当提到的是，一些示例性实施例被描述成作为流程图描绘的处理或方法。虽然流程图将各项操作(或步骤)描述成顺序的处理，但是其中的许多操作可以被并行地、并发地或者同时实施。此外，各项操作的顺序可以被重新安排。当其操作完成时所述处理可以被终止，但是还可以具有未包括在附图中的附加步骤。所述处理可以对应于方法、函数、规程、子例程、子程序等等。
需要说明的是，本申请实施例中，“示例性的”或者“例如”等词用于表示作例子、例证或说明。本申请实施例中被描述为“示例性的”或者“例如”的任何实施例或设计方案不应被解释为比其它实施例或设计方案更优选或更具优势。确切而言，使用“示例性的”或者“例如”等词旨在以举例方式呈现相关概念。
在本申请的描述中，除非另有说明，“多个”的含义是指两个或两个以上。
近年来，人工智能相关产业不断发展，依托于大数据信息时代，深度学习方法不断被运用到各个领域，卷积神经网络的飞速发展推动了计算机视觉的不断进步，基于深度学习的目标检测技术已经在各个产业领域不断落地。深度学习模型包括一般目标检测和特定目标检测，目标检测器可以利用深度学习网络作为主干和检测网络，从输入图像或视频中提取特征然后进行分类和定位，以实现目标检测。主流目标检测器通常可以分为两类，一类是两级检测器，例如Faster R-CNN；另一类是一级检测器，例如YOLO和SSD。一级检测器具有较高的推理速度，两级检测器具有较高的定位和目标识别精度。
对于火焰检测问题，可以在潜在的容易发生火灾的地点部署摄像头，以全天候智能监测的方式进行火焰安全检测，将火灾消灭在萌芽之中。
但是，火焰检测所需的实时性和准确性较高，相关技术的目标检测器不能满足火焰检测的需求，不能直接应用于火焰检测。
因此，本申请提出一种火焰检测方法，以实现高精度和高实时性的火焰检测。
下面，将结合各实施例对本申请所提出的火焰检测方法进行描述。
实施例一
图1为本申请实施例一提供的一种火焰检测方法的流程图，本实施例可适用于基于监控视频实现快速且精确的火焰检测的情况，该方法可以由火焰检测装置来执行，如图1所示，包括如下步骤：
步骤110、将待测视频输入预先训练好的火焰检测模型，以使所述火焰检测模型根据所述待测视频确定待测图像，并对所述待测图像进行分类，得到的输出结果为所述待测图像的类别信息。
在一实施例中，火焰检测模型可以包括筛选模块和检测模块；筛选模块可以基于相关技术的视频帧筛选方法对待测视频所包含的视频帧进行筛选，以根据待测视频确定待测图像；检测模块可以基于深度学习的目标检测算法对待测图像进行分类，以确定待测图像的类别信息。
示例性地，首先可以从容易发生火灾的设施地点的摄像头获取监控视频，并将该监控视频确定为待测视频，然后将待测视频作为输入信息输入火焰检测模型。火焰检测模型首先可以基于筛选模块对待测视频所包含的多个视频帧进行筛选，将可能包含火焰的视频帧确定为待测图像；然后将待测图像输入检测模块中，检测模块可以对待测图像进行分类，以实现对待测图像的分类和定位。当然，火焰检测模型的输出信息可以为待测图像的类别信息，火焰检测模型实现了对待测视频的火焰检测。
火焰检测模型首先基于筛选模块对待测视频的多个视频帧进行了筛选，然后将通过筛选得到的待测图像输入检测模块，实现了对视频帧的过滤，将可能存在火焰的视频帧全部筛查，降低了输入检测模块的数据量，提高了火焰检测的检测效率。检测模块可以基于输入的待测图像进行火焰检测，得到待测视频对应的火焰检测结果。
本申请实施例中，预先训练好的火焰检测模型可以将相关技术的视频帧筛选方法与深度学习的目标检测方法进行结合，在对视频帧进行目标检测之前，先用相关技术的视频帧过滤方法对待测视频所包含的视频帧进行视频帧过滤，将可能存在火焰的视频帧全部筛查，并将可能存在火焰的视频帧确定为待测图像，降低火焰检测的数据量，提升火焰检测的检测效率。当然，还可以继续对待测图像进行目标检测，得到待测视频对应的火焰检测结果，提升了火焰检测的实时性和准确性。
步骤120、根据所述类别信息确定火焰检测结果。
在一实施例中，类别信息可以包括：待测图像包含火焰以及待测图像不包含火焰。
示例性地，如果类别信息为待测图像包含火焰，则可以确定火焰检测结果为获取到待测视频的图像获取装置的所处位置存在火焰；如果类别信息为待测 图像不包含火焰，则确定火焰检测结果为获取到待测视频的图像获取装置的所处位置不存在火焰。
本申请实施例中，根据火焰检测模型确定的待测图像的类别信息可以确定待测图像的火焰检测结果，进而可以确定待测图像所属待测视频的火焰检测结果，当然，还可以确定获取待测视频的图像获取装置所处位置的火焰检测结果，以实现对易发生火灾位置的监控，降低火灾发生的概率。
本申请实施例一提供的一种火焰检测方法，包括：将待测视频输入预先训练好的火焰检测模型，以使所述火焰检测模型根据所述待测视频确定待测图像，并对所述待测图像进行分类，得到的输出结果为所述待测图像的类别信息；根据所述类别信息确定火焰检测结果。上述技术方案，预先训练好的火焰检测模型可以根据待测视频确定待测图像，降低目标检测的数据量，还可以基于目标检测对待测图像进行分类，并根据分类结果确定火焰检测结果，提高了对于火焰的感知能力，也提高了火焰检测的性能，同时提高了火焰检测的实时性和可靠性。
实施例二
图2为本申请实施例二提供的一种火焰检测方法的流程图，本实施例是在上述实施例的基础上进行改动。如图2所示，在本实施例中，该方法还可以包括：
步骤210、基于公共图像集预训练预设目标检测模型得到预训练模型。
例如，公共图像集可以为Image Net数据集，Image Net数据集是一个计算机视觉数据集，包含多种图像以及多种图像对应的图像类别。
示例性地，可以在公共图像集上预训练预设目标检测模型得到预训练模型，在实际应用中，可以在Image Net数据集上预训练预设目标检测模型，得到预训练模型。
当然，由于Image Net数据集中火焰图像较少，因此，经过Image Net数据集预训练预设目标检测模型得到的预训练模型对于火焰检测的精确度较低，需要将预训练模型进行迁移，在火焰图像集中进行再一次训练。
一种实施方式中，在训练预设目标检测模型得到预训练模型的同时，可以确定第一参数，该第一参数可以用于确定火焰检测模型。
在对预训练模型进行迁移时，可以冻结第一参数，并对预训练模型的其他参数重新训练。冻结的第一参数可以用于确定火焰检测模型。当然，对于第一参数的选取，可以根据实际需求进行确定。
本申请实施例中，由于火焰图像的样本较少，首先可以在Image Net数据集上对预设目标检测模型进行预训练，以得到预训练模型，解决了训练集图像较少导致的火焰检测模型难以进行准确训练的问题。
步骤220、对火焰图像集所包含的多个火焰图像进行特征匹配，确定所述多个火焰图像的真实标签。
示例性地，首先，可以通过特征提取网络提取火焰图像集所包含多个火焰图像的图像特征；其次，可以基于自由锚框(Anchor-free)机制快速生成图像特征对应的火焰图像的锚框，并将锚框用于模型训练，同时减少模型的参数量；进而，可以基于样本高效匹配方案(如SimOTA方案)对锚框和标签进行特征匹配。此处的标签可以为真实标签，即火焰。
需要说明的是，Anchor-free机制相比于基于锚框的模型显著减少了需要启发式调整的设计参数的数量和涉及的许多技巧(例如，锚定群集、网格敏感)，使得火焰检测模型的训练和解码阶段变得相当简单。SimOTA方案自动为火焰目标设定不同的正样本数量，提高了火焰检测的效率。通过代价矩阵自适应的找到当前特征点应该去拟合的火焰真实框，重合度越高越需要拟合，分类越准越需要拟合，在一定半径内越需要拟合。
图3为本申请实施例二提供的一种火焰检测方法中火焰检测模型的示意图，如图3所示，一种实施方式中，所述火焰检测模型包括检测模块，所述检测模块包括：特征提取单元、自由锚框(Anchor-free)单元和样本高效匹配(SimOTA)单元，相应地，步骤220可以包括：
基于所述特征提取单元确定所述多个火焰图像的特征信息；通过所述Anchor-free单元生成所述特征信息对应的所述火焰图像的锚框；采用所述SimOTA单元对所述锚框进行特征匹配，以确定所述锚框对应的所述火焰图像的真实标签。
图4为本申请实施例二提供的一种火焰检测方法中火焰检测模型的训练示意图，如图4所示，首先，可以通过特征提取单元确定火焰图像集所包含多个火焰图像的特征信息；其次，可以通过Anchor-free单元快速生成特征信息对应的火焰图像的锚框；进而，可以采用SimOTA单元对锚框进行特征匹配，以确定锚框对应的火焰图像的真实标签。此处的真实标签可以火焰。
需要说明的是，YOLOX网络(一种高性能检测器)为了达到最优的检测性能，需要在训练前进行聚类分析，确定一组最优的锚点(anchors)。那些聚集的锚点是特定域的并且不太普遍，而且增加了检测头的复杂性，以及每个图像的预测数量。在某些边缘人工智能(Artificial Intelligence，AI)系统上，在设备之间(例如，从神经网络处理器(Neural network Processing Unit，NPU)到中央处理器(Central Processing Unit，CPU))移动如此大量的预测可能会成为整体延迟方面的潜在瓶颈。因此，可以采用Anchor-free单元确定锚框。在实际应用中，将YOLO网络切换为Anchor-free单元的方式非常简单，可以将每个位置的预测从3个减少到1个，并使它们直接预测四个值，即网格左上角的两个偏移量，以及预测框的高度和宽度。将每个对象的中心位置分配为正样本，并预先定义一个刻度范围，以指定每个对象的特征金字塔网络(Feature Pyramid Network，FPN) 级别。Anchor-free单元的应用减少了火焰检测模型的参数和计算力，但获得了更好的性能。
还需要说明的是，SimOTA单元首先可以计算锚框和真实标签的成对匹配度，由每个预测-真值(groundtruth)对的损失表示。例如，在SimOTA单元中，groundtruth和预测之间的损失为其中，λ为平衡系数，为类别损失，为回归损失。对于真实标签，可以选择固定中心区域内损失最低的设定个预测作为它的正样本。最后，这些正预测的相应网格被指定为正值，而其余的网格是负值。在实际应用中，正样本数量的确定可以根据groundtruth确定。SimOTA单元不仅减少了火焰检测模型的训练时间，而且避免了额外的超参数求解。SimOTA单元将火焰检测模型的平均精准度(Average Precision，AP)从45.0％提升到47.3％，提升了火焰检测模型的准确性。
本申请实施例中，在对火焰检测模型进行训练的过程中，通过Anchor-free单元解决基于锚框机制所产生的大量冗余参数问题，提高模型训练效率；采用SimOTA单元缓解极端长宽比的物体的检测效果差的问题，以及极端大小目标正样本不均衡的问题，同时缓解旋转物体检测效果不好的问题，提高模型训练效率。
步骤230、将所述火焰图像集所包含的所述多个火焰图像以及所述多个火焰图像的真实标签作为训练数据对所述预训练模型进行网络训练，并计算损失函数。
示例性地，在将火焰图像集所包含的多个火焰图像作为输入信息输入预训练模型后，预训练模型可以进行目标的回归和识别，以确定火焰图像的训练类别，比较预训练模型确定的训练类别和真实标签确定的真实类别，并根据比较结果计算损失函数。
在实际应用中，损失函数可以根据实际需求进行选取。
步骤240、基于反向传播算法进行网络优化，直至所述损失函数收敛，得到所述火焰检测模型。
示例性地，在模型训练过程中，可以基于反向传播算法进行网络优化，直至损失函数收敛。损失函数收敛，表明预训练模型训练完成，进而可以将损失函数收敛的预训练模型确定为火焰检测模型。
基于反向传播算法进行网络优化，在损失函数收敛时，确定第二参数，该第二参数也可以用于确定火焰检测模型。进而，可以根据前述确定预训练模型时确定的第一参数，以及损失函数收敛时确定的第二参数，可以确定火焰检测模型。
本申请实施例中，针对火焰图像较少的问题，首先在Image Net数据集训练预设目标检测模型得到预训练模型，再引入迁移学习技术，将预训练模型进行迁移，迁移的同时冻结部分参数。在火焰图像集上对预训练模型重新训练得到 火焰检测模型。为了加快产品落地速度，可以采用SimOTA单元确定火焰图像集所包含多个火焰图像的真实标签，以在更短的时间内完成火焰检测模型的训练。在模型优化方面，现有目标检测算法，大部分都是基于锚框的机制，这样的设计就需要大量的参数，需要手动设定的参数包括：尺度和长宽比。而针对不同检测目标，参数中的尺度和长宽比比较难设计。采用Anchor-free单元确定锚框，可以很大程度上避免以上问题，可以提供更大更灵活的解空间，并且摆脱了使用锚框而带来的计算量，提升模型训练的精度。
步骤250、将待测视频输入预先训练好的火焰检测模型，以使所述火焰检测模型根据所述待测视频确定待测图像，并对所述待测图像进行分类，得到的输出结果为所述待测图像的类别信息。
如图3所示，一种实施方式中，所述火焰检测模型还包括筛选模块，所述检测模块还包括主体单元，相应地，所述火焰检测模型根据所述待测视频确定待测图像，并对所述待测图像进行分类，得到的输出结果为所述待测图像的类别信息，包括：
通过所述筛选模块基于边缘检测算法对所述待测视频所包含的视频帧进行筛选，得到所述待测图像；基于所述检测模块中的所述Anchor-free单元确定所述待测图像的锚框，将所述锚框输入所述主体单元，以使得所述主体单元根据所述锚框确定所述待测图像的类别信息。
例如，主体单元可以为YOLOX的主体检测网络。
示例性地，筛选模块可以基于更加快速的筛选方法对待测视频所包含的视频帧进行大量筛选，将可能存在火焰的图像确定为待测图像。进而可以基于检测模块所包含的特征提取单元确定待测图像的特征信息，并基于检测模块所包含的Anchor-free单元确定特征信息对应的待测图像的锚框。当然，可以将锚框输入主体检测网络，主体检测网络可以进行目标的回归和识别，以确定待测图像的类别信息。
一种实施方式中，所述通过筛选模块基于边缘检测算法对所述待测视频所包含的视频帧进行筛选，得到所述待测图像，包括：
通过所述筛选模块基于边缘检测算法对所述待测视频所包含的所述视频帧进行图像分类，得到分类结果；将所述分类结果为包含火焰的所述视频帧确定为所述待测图像。
一种实施方式中，所述边缘检测算法包括动态阈值分割算法、均值滤波算法和阈值分割算法，相应地，通过所述筛选模块基于边缘检测算法对所述待测视频所包含的所述多个视频帧进行图像分类，得到分类结果，包括：
基于所述动态阈值分割算法、所述均值滤波算法和所述阈值分割算法，分别确定所述视频帧的灰度值；对基于所述动态阈值分割算法确定的所述视频帧的所述灰度值、基于所述均值滤波算法确定的所述视频帧的所述灰度值以及基 于所述阈值分割算确定的所述视频的所述灰度值进行加权求和，以得到所述视频帧的目标灰度值；根据所述目标灰度值确定所述视频帧的所述分类结果。
本申请实施例中，火焰检测模型可以将相关技术的视频帧筛选方法与深度学习的目标检测方法进行结合，在对视频帧进行目标检测之前，先用相关技术的视频帧过滤方法对待测视频所包含的视频帧进行视频帧过滤，例如可以基于边缘检测算法对视频帧进行图像分类，并将分类结果为可能存在火焰的视频帧全部筛查，进而将可能存在火焰的视频帧确定为待测图像，降低火焰检测的数据量，提升火焰检测的检测效率。当然，还可以继续对待测图像进行目标检测，得到待测视频对应的火焰检测结果，提升了火焰检测的实时性和准确性。
步骤260、根据所述类别信息确定火焰检测结果。
一种实施方式中，步骤260可以包括：
如果所述类别信息为所述待测图像包含火焰，则确定所述火焰检测结果为获取到所述待测视频的图像获取装置的所处位置存在火焰；如果所述类别信息为所述待测图像不包含火焰，则确定所述火焰检测结果为获取到所述待测视频的图像获取装置的所处位置不存在火焰。
本申请实施例中，根据火焰检测模型确定的待测图像的类别信息可以确定待测图像的火焰检测结果，进而可以确定待测图像所属待测视频的火焰检测结果，当然，还可以确定获取待测视频的图像获取装置所处位置的火焰检测结果，以实现对易发生火灾位置的监控，降低火灾发生的概率。
本申请实施例二提供的一种火焰检测方法，包括：基于公共图像集预训练预设目标检测模型得到预训练模型；对火焰图像集所包含的多个火焰图像进行特征匹配，确定所述多个火焰图像的真实标签；将所述火焰图像集所包含的所述多个火焰图像以及所述多个火焰图像的真实标签作为训练数据对所述预训练模型进行网络训练，并计算损失函数；基于反向传播算法进行网络优化，直至所述损失函数收敛，得到所述火焰检测模型；将待测视频输入预先训练好的火焰检测模型，以使所述火焰检测模型根据所述待测视频确定待测图像，并对所述待测图像进行分类，得到的输出结果为所述待测图像的类别信息；根据所述类别信息确定火焰检测结果。上述技术方案，首先在Image Net数据集训练预设目标检测模型得到预训练模型，再引入迁移学习技术，将预训练模型进行迁移，在火焰图像集上对预训练模型重新训练得到火焰检测模型。例如可以采用SimOTA单元确定火焰图像集所包含多个火焰图像的真实标签，以在更短的时间内完成火焰检测模型的训练。采用Anchor-free单元确定锚框，可以提供更大更灵活的解空间，并且摆脱了使用锚框而带来的计算量，提升模型训练的精度。当然，训练好的火焰检测模型可以将相关技术的视频帧筛选方法与深度学习的目标检测方法进行结合，在对视频帧进行目标检测之前，先用相关技术的视频帧过滤方法对待测视频所包含的视频帧进行视频帧过滤，将可能存在火焰的视频帧确定为待测图像，降低火焰检测的数据量，提升火焰检测的检测效率。提高 了对于火焰的感知能力，也提高了火焰检测的性能，同时提高了火焰检测的实时性和可靠性。
实施例三
图5为本申请实施例三提供的一种火焰检测装置的结构示意图，该装置可以适用于基于监控视频实现快速且精确的火焰检测的情况，提高火焰检测的实时性和准确性。该装置可以通过软件和/或硬件实现，并一般集成在计算机设备中。
如图5所示，该装置包括：
执行模块510，设置为将待测视频输入预先训练好的火焰检测模型，以使所述火焰检测模型根据所述待测视频确定待测图像，并对所述待测图像进行分类，得到的输出结果为所述待测图像的类别信息；
确定模块520，设置为根据所述类别信息确定火焰检测结果。
本实施例提供的火焰检测装置，通过将待测视频输入预先训练好的火焰检测模型，以使所述火焰检测模型根据所述待测视频确定待测图像，并对所述待测图像进行分类，得到的输出结果为所述待测图像的类别信息；根据所述类别信息确定火焰检测结果。上述技术方案，预先训练好的火焰检测模型可以根据待测视频确定待测图像，降低目标检测的数据量，还可以基于目标检测对待测图像进行分类，并根据分类结果确定火焰检测结果，提高了对于火焰的感知能力，也提高了火焰检测的性能，同时提高了火焰检测的实时性和可靠性。
在上述实施例的基础上，所述火焰检测模型通过如下方式得到：
基于公共图像集预训练预设目标检测模型得到预训练模型；
对火焰图像集所包含的多个火焰图像进行特征匹配，确定所述多个火焰图像的真实标签；
将所述火焰图像集所包含的所述多个火焰图像以及所述多个火焰图像的真实标签作为训练数据对所述预训练模型进行网络训练，并计算损失函数；
基于反向传播算法进行网络优化，直至所述损失函数收敛，得到所述火焰检测模型。
在上述实施例的基础上，所述火焰检测模型包括检测模块，所述检测模块包括：特征提取单元、自由锚框单元和样本高效匹配单元，
对火焰图像集所包含的多个火焰图像进行特征匹配，确定所述多个火焰图像的真实标签，包括：
基于所述特征提取单元确定所述多个火焰图像的特征信息；
通过所述自由锚框单元生成所述特征信息对应的所述火焰图像的锚框；
采用所述样本高效匹配单元对所述锚框进行特征匹配，以确定所述锚框对 应的所述火焰图像的真实标签。
在上述实施例的基础上，所述火焰检测模型还包括筛选模块，所述检测模块还包括主体单元，
所述火焰检测模型根据所述待测视频确定待测图像，并对所述待测图像进行分类，得到的输出结果为所述待测图像的类别信息，包括：
通过所述筛选模块基于边缘检测算法对所述待测视频所包含的视频帧进行筛选，得到所述待测图像；
基于所述检测模块中的所述自由锚框单元确定所述待测图像的锚框，并将所述锚框输入所述主体单元，以使得所述主体单元根据所述锚框确定所述待测图像的类别信息。
在上述实施例的基础上，所述通过筛选模块基于边缘检测算法对所述待测视频所包含的视频帧进行筛选，得到所述待测图像，包括：
通过所述筛选模块基于边缘检测算法对所述待测视频所包含的所述视频帧进行图像分类，得到分类结果；
将所述分类结果为包含火焰的所述视频帧确定为所述待测图像。
在上述实施例的基础上，所述边缘检测算法包括动态阈值分割算法、均值滤波算法和阈值分割算法，
通过所述筛选模块基于边缘检测算法对所述待测视频所包含的所述视频帧进行图像分类，得到分类结果，包括：
基于所述动态阈值分割算法、所述均值滤波算法和所述阈值分割算法，分别确定所述视频帧的灰度值；
对基于所述动态阈值分割算法确定的所述视频帧的所述灰度值、基于所述均值滤波算法确定的所述视频帧的所述灰度值以及基于所述阈值分割算确定的所述视频的所述灰度值进行加权求和，以得到所述视频帧的目标灰度值；
根据所述目标灰度值确定所述视频帧的所述分类结果。
在上述实施例的基础上，确定模块520，设置为通过以下方式根据目标灰度值确定视频帧的分类结果：
响应于所述类别信息为所述待测图像包含火焰，确定所述火焰检测结果为获取到所述待测视频的图像获取装置的所处位置存在火焰；
响应于所述类别信息为所述待测图像不包含火焰，确定所述火焰检测结果为获取到所述待测视频的图像获取装置的所处位置不存在火焰。
本申请实施例所提供的火焰检测装置可执行本申请任意实施例所提供的火焰检测方法，具备执行方法相应的功能模块和有益效果。
实施例四
图6为本申请实施例四提供的一种计算机设备的结构示意图。图6示出了适于用来实现本申请实施方式的示例性计算机设备6的框图。图6显示的计算机设备6仅仅是一个示例。
如图6所示，计算机设备6以通用计算设备的形式表现。计算机设备6的组件可以包括：至少一个处理器或者处理单元16，系统存储器28，连接不同系统组件(包括系统存储器28和处理单元16)的总线18。
总线18表示几类总线结构中的一种或多种，包括存储器总线或者存储器控制器，外围总线，图形加速端口，处理器或者使用多种总线结构中的任意总线结构的局域总线。举例来说，这些体系结构包括工业标准体系结构(International Standards Association，ISA)总线，微通道体系结构(Micro Channel Architecture，MCA)总线，增强型ISA总线、视频电子标准协会(Video Electronics Standards Association，VESA)局域总线以及外围组件互连(Peripheral Component Interconnect，PCI)总线。
计算机设备6典型地包括多种计算机系统可读介质。这些介质可以是任何能够被计算机设备6访问的可用介质，包括易失性和非易失性介质，可移动的和不可移动的介质。
系统存储器28可以包括易失性存储器形式的计算机系统可读介质，例如随机存取存储器(Random Access Memory，RAM)30和/或高速缓存存储器32。计算机设备6可以包括其它可移动/不可移动的、易失性/非易失性计算机系统存储介质。仅作为举例，存储系统34可以用于读写不可移动的、非易失性磁介质(通常称为“硬盘驱动器”)。可以提供用于对可移动非易失性磁盘(例如“软盘”)读写的磁盘驱动器，以及对可移动非易失性光盘(例如便携式紧凑磁盘只读存储器(Compact Disc-Read Only Memory，CD-ROM)，只读数字激光视盘机(Digital Video Disc-Read Only Memory，DVD-ROM)或者其它光介质)读写的光盘驱动器。在这些情况下，每个驱动器可以通过至少一个数据介质接口与总线18相连。系统存储器28可以包括至少一个程序产品，该程序产品具有一组(例如至少一个)程序模块，这些程序模块被配置以执行本申请各实施例的功能。
具有一组(至少一个)程序模块42的程序/实用工具40，可以存储在例如系统存储器28中，这样的程序模块42包括操作系统、至少一个应用程序、其它程序模块以及程序数据，这些示例中的每一个或某种组合中可能包括网络环境的实现。程序模块42通常执行本申请所描述的实施例中的功能和/或方法。
计算机设备6也可以与至少一个外部设备14(例如键盘、指向设备、显示器24等)通信，还可与至少一个使得用户能与该计算机设备6交互的设备通信，和/或与使得该计算机设备6能与一个或多个其它计算设备进行通信的任何设备(例如网卡，调制解调器等等)通信。这种通信可以通过输入/输出(Input/Output，I/O)接口22进行。并且，计算机设备6还可以通过网络适配器20与至少一个网络 (例如局域网(Local Area Network，LAN)，广域网(Wide Area Network，WAN)和/或公共网络，例如因特网)通信。如图6所示，网络适配器20通过总线18与计算机设备6的其它模块通信。应当明白，可以结合计算机设备6使用其它硬件和/或软件模块，包括：微代码、设备驱动器、冗余处理单元、外部磁盘驱动阵列、磁盘阵列(Redundant Arrays of Independent Disks，RAID)系统、磁带驱动器以及数据备份存储系统等。
处理单元16通过运行存储在系统存储器28中的程序，从而执行各种功能应用以及页面显示，例如实现本申请实施例所提供的火焰检测方法，该方法包括：
将待测视频输入预先训练好的火焰检测模型，以使所述火焰检测模型根据所述待测视频确定待测图像，并对所述待测图像进行分类，得到的输出结果为所述待测图像的类别信息；
根据所述类别信息确定火焰检测结果。
当然，本领域技术人员可以理解，处理器还可以实现本申请任意实施例所提供的火焰检测方法的技术方案。
实施例五
本申请实施例五提供了一种计算机可读存储介质，计算机可读存储介质上存储有计算机程序，该计算机程序被处理器执行时实现例如本发实施例所提供的火焰检测方法，该方法包括：
将待测视频输入预先训练好的火焰检测模型，以使所述火焰检测模型根据所述待测视频确定待测图像，并对所述待测图像进行分类，得到的输出结果为所述待测图像的类别信息；
根据所述类别信息确定火焰检测结果。
本申请实施例的计算机存储介质，可以采用一个或多个计算机可读的介质的组合。计算机可读介质可以是计算机可读信号介质或者计算机可读存储介质。计算机可读存储介质例如可以是：电、磁、光、电磁、红外线、或半导体的系统、装置或器件，或者以上的组合。计算机可读存储介质的示例包括：具有至少一个导线的电连接、便携式计算机磁盘、硬盘、随机存取存储器(RAM)、只读存储器(Read Only Memory，ROM)、可擦式可编程只读存储器(如电子可编程只读存储器(Electronic Programable Read Only Memory，EPROM)或闪存)、光纤、便携式紧凑磁盘只读存储器(CD-ROM)、光存储器件、磁存储器件、或者上述的合适的组合。在本文件中，计算机可读存储介质可以是包含或存储程序的有形介质，该程序可以被指令执行系统、装置或者器件使用或者与其结合使用。
计算机可读的信号介质可以包括在基带中或者作为载波一部分传播的数据信号，其中承载了计算机可读的程序代码。这种传播的数据信号可以采用多种 形式，包括电磁信号、光信号或上述的合适的组合。计算机可读的信号介质还可以是计算机可读存储介质以外的任何计算机可读介质，该计算机可读介质可以发送、传播或者传输用于由指令执行系统、装置或者器件使用或者与其结合使用的程序。
计算机可读介质上包含的程序代码可以用任何适当的介质传输，包括但不限于：无线、电线、光缆、射频(Radio Frequency，RF)等，或者上述的合适的组合。
可以以至少一种程序设计语言或其组合来编写用于执行本申请操作的计算机程序代码，程序设计语言包括面向对象的程序设计语言，诸如Java、Smalltalk、C++，还包括常规的过程式程序设计语言，诸如“C”语言或类似的程序设计语言。程序代码可以完全地在用户计算机上执行、部分地在用户计算机上执行、作为一个独立的软件包执行、部分在用户计算机上部分在远程计算机上执行、或者完全在远程计算机或服务器上执行。在涉及远程计算机的情形中，远程计算机可以通过任意种类的网络，包括局域网(LAN)或广域网(WAN)，连接到用户计算机，或者，可以连接到外部计算机(例如利用因特网服务提供商来通过因特网连接)。
本领域普通技术人员应该明白，上述的本申请的各模块或各步骤可以用通用的计算装置来实现，它们可以集中在单个计算装置上，或者分布在多个计算装置所组成的网络上，可选地，它们可以用计算机装置可执行的程序代码来实现，从而可以将它们存储在存储装置中由计算装置来执行，或者将它们分别制作成各个集成电路模块，或者将它们中的多个模块或步骤制作成单个集成电路模块来实现。这样，本申请不限制于任何特定的硬件和软件的结合。

Claims (10)

  1. 一种火焰检测方法，包括：
    将待测视频输入预先训练好的火焰检测模型，以使所述火焰检测模型根据所述待测视频确定待测图像，并对所述待测图像进行分类，得到的输出结果为所述待测图像的类别信息；
    根据所述类别信息确定火焰检测结果。
  2. 根据权利要求1所述的火焰检测方法，其中，所述火焰检测模型通过如下方式得到：
    基于公共图像集预训练预设目标检测模型得到预训练模型；
    对火焰图像集所包含的多个火焰图像进行特征匹配，确定所述多个火焰图像的真实标签；
    将所述火焰图像集所包含的所述多个火焰图像以及所述多个火焰图像的真实标签作为训练数据对所述预训练模型进行网络训练，并计算损失函数；
    基于反向传播算法进行网络优化，直至所述损失函数收敛，得到所述火焰检测模型。
  3. 根据权利要求2所述的火焰检测方法，其中，所述火焰检测模型包括检测模块，所述检测模块包括：特征提取单元、自由锚框单元和样本高效匹配单元，
    对火焰图像集所包含的多个火焰图像进行特征匹配，确定所述多个火焰图像的真实标签，包括：
    基于所述特征提取单元确定所述多个火焰图像的特征信息；
    通过所述自由锚框单元生成所述特征信息对应的所述火焰图像的锚框；
    采用所述样本高效匹配单元对所述锚框进行特征匹配，以确定所述锚框对应的所述火焰图像的真实标签。
  4. 根据权利要求3所述的火焰检测方法，其中，所述火焰检测模型还包括筛选模块，所述检测模块还包括主体单元，
    所述火焰检测模型根据所述待测视频确定待测图像，并对所述待测图像进行分类，得到的输出结果为所述待测图像的类别信息，包括：
    通过所述筛选模块基于边缘检测算法对所述待测视频所包含的视频帧进行筛选，得到所述待测图像；
    基于所述检测模块中的所述自由锚框单元确定所述待测图像的锚框，并将所述锚框输入所述主体单元，以使得所述主体单元根据所述锚框确定所述待测图像的类别信息。
  5. 根据权利要求4所述的火焰检测方法，其中，所述通过筛选模块基于边缘检测算法对所述待测视频所包含的视频帧进行筛选，得到所述待测图像，包括：
    通过所述筛选模块基于边缘检测算法对所述待测视频所包含的所述视频帧进行图像分类，得到分类结果；
    将所述分类结果为包含火焰的所述视频帧确定为所述待测图像。
  6. 根据权利要求5所述的火焰检测方法，其中，所述边缘检测算法包括动态阈值分割算法、均值滤波算法和阈值分割算法，
    通过所述筛选模块基于边缘检测算法对所述待测视频所包含的所述视频帧进行图像分类，得到分类结果，包括：
    基于所述动态阈值分割算法、所述均值滤波算法和所述阈值分割算法，分别确定所述视频帧的灰度值；
    对基于所述动态阈值分割算法确定的所述视频帧的所述灰度值、基于所述均值滤波算法确定的所述视频帧的所述灰度值以及基于所述阈值分割算确定的所述视频的所述灰度值进行加权求和，以得到所述视频帧的目标灰度值；
    根据所述目标灰度值确定所述视频帧的所述分类结果。
  7. 根据权利要求1所述的火焰检测方法，其中，根据所述类别信息确定火焰检测结果，包括：
    响应于所述类别信息为所述待测图像包含火焰，确定所述火焰检测结果为：获取到所述待测视频的图像获取装置的所处位置存在火焰；
    响应于所述类别信息为所述待测图像不包含火焰，确定所述火焰检测结果为：获取到所述待测视频的图像获取装置的所处位置不存在火焰。
  8. 一种火焰检测装置，包括：
    执行模块，被设置为将待测视频输入预先训练好的火焰检测模型，以使所述火焰检测模型根据所述待测视频确定待测图像，并对所述待测图像进行分类，得到的输出结果为所述待测图像的类别信息；
    确定模块，被设置为根据所述类别信息确定火焰检测结果。
  9. 一种计算机设备，所述设备包括：
    处理器；以及与所述处理器通信连接的存储器；
    其中，所述存储器存储有可被所述处理器执行的计算机程序，在所述计算机程序被所述处理器执行时，所述处理器能够执行权利要求1-7中任一项所述的 火焰检测方法。
  10. 一种包含计算机可执行指令的存储介质，所述计算机可执行指令在由计算机处理器执行时用于执行如权利要求1-7中任一所述的火焰检测方法。
PCT/CN2023/073529 2022-01-27 2023-01-28 火焰检测方法、装置、设备和存储介质 WO2023143498A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202210097095.8A CN114120208A (zh) 2022-01-27 2022-01-27 一种火焰检测方法、装置、设备和存储介质
CN202210097095.8 2022-01-27

Publications (1)

Publication Number Publication Date
WO2023143498A1 true WO2023143498A1 (zh) 2023-08-03

Family

ID=80361284

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2023/073529 WO2023143498A1 (zh) 2022-01-27 2023-01-28 火焰检测方法、装置、设备和存储介质

Country Status (2)

Country Link
CN (1) CN114120208A (zh)
WO (1) WO2023143498A1 (zh)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114120208A (zh) * 2022-01-27 2022-03-01 青岛海尔工业智能研究院有限公司 一种火焰检测方法、装置、设备和存储介质

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20170286774A1 (en) * 2016-04-04 2017-10-05 Xerox Corporation Deep data association for online multi-class multi-object tracking
CN108052865A (zh) * 2017-07-06 2018-05-18 同济大学 一种基于卷积神经网络和支持向量机的火焰检测方法
CN109815863A (zh) * 2019-01-11 2019-05-28 北京邮电大学 基于深度学习和图像识别的烟火检测方法和系统
CN114120208A (zh) * 2022-01-27 2022-03-01 青岛海尔工业智能研究院有限公司 一种火焰检测方法、装置、设备和存储介质

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102034110B (zh) * 2010-12-09 2013-02-27 湘潭乐星电气有限公司 一种火焰检测方法
CN108038867A (zh) * 2017-12-22 2018-05-15 湖南源信光电科技股份有限公司 基于多特征融合和立体视觉的火焰检测和定位方法
CN109377713B (zh) * 2018-09-26 2021-03-16 石化盈科信息技术有限责任公司 一种火灾预警方法及系统
CN112052797A (zh) * 2020-09-07 2020-12-08 合肥科大立安安全技术有限责任公司 基于MaskRCNN的视频火灾识别方法及系统

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20170286774A1 (en) * 2016-04-04 2017-10-05 Xerox Corporation Deep data association for online multi-class multi-object tracking
CN108052865A (zh) * 2017-07-06 2018-05-18 同济大学 一种基于卷积神经网络和支持向量机的火焰检测方法
CN109815863A (zh) * 2019-01-11 2019-05-28 北京邮电大学 基于深度学习和图像识别的烟火检测方法和系统
CN114120208A (zh) * 2022-01-27 2022-03-01 青岛海尔工业智能研究院有限公司 一种火焰检测方法、装置、设备和存储介质

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
JIAHAO SUN, LI DAN: "Fire alarm system based on YOLOX", PRACTICAL ELECTRONICS, vol. 22, 15 November 2021 (2021-11-15), pages 49 - 51, XP093081903, DOI: 10.16589/j.cnki.cn11-3571/tn.2021.22.017 *

Also Published As

Publication number Publication date
CN114120208A (zh) 2022-03-01

Similar Documents

Publication Publication Date Title
JP7331171B2 (ja) 画像認識モデルをトレーニングするための方法および装置、画像を認識するための方法および装置、電子機器、記憶媒体、並びにコンピュータプログラム
CN111723786B (zh) 一种基于单模型预测的安全帽佩戴检测方法及装置
WO2020164282A1 (zh) 基于yolo的图像目标识别方法、装置、电子设备和存储介质
CN112132119B (zh) 客流统计方法、装置、电子设备和存储介质
CN112733749A (zh) 融合注意力机制的实时行人检测方法
CN110875912A (zh) 一种基于深度学习的网络入侵检测方法、装置和存储介质
CN110991311A (zh) 一种基于密集连接深度网络的目标检测方法
WO2021103868A1 (zh) 一种行人结构化方法、装置、设备和存储介质
CN110968734B (zh) 一种基于深度度量学习的行人重识别方法及装置
Chen et al. An improved Yolov3 based on dual path network for cherry tomatoes detection
WO2023143498A1 (zh) 火焰检测方法、装置、设备和存储介质
WO2021175278A1 (zh) 一种模型更新方法以及相关装置
CN110458022A (zh) 一种基于域适应的可自主学习目标检测方法
Racah et al. Semi-supervised detection of extreme weather events in large climate datasets
Amisse et al. Fine-tuning deep learning models for pedestrian detection
CN116579616A (zh) 一种基于深度学习的风险识别方法
Shankar et al. A framework to enhance object detection performance by using YOLO algorithm
CN115082740A (zh) 目标检测模型训练方法、目标检测方法、装置、电子设备
CN106529455A (zh) 一种基于SoC FPGA的快速人体姿态识别方法
CN111222534A (zh) 一种基于双向特征融合和更平衡l1损失的单发多框检测器优化方法
CN113569657A (zh) 一种行人重识别方法、装置、设备及存储介质
CN112907138B (zh) 一种从局部到整体感知的电网场景预警分类方法及系统
WO2023232031A1 (zh) 一种神经网络模型的训练方法、装置、电子设备及介质
CN115482436B (zh) 图像筛选模型的训练方法、装置以及图像筛选方法
CN116630753A (zh) 一种基于对比学习的多尺度小样本目标检测方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 23746375

Country of ref document: EP

Kind code of ref document: A1