WO2022247344A1 - 视频识别模型训练方法、装置、设备以及存储介质 - Google Patents

视频识别模型训练方法、装置、设备以及存储介质 Download PDF

Info

Publication number
WO2022247344A1
WO2022247344A1 PCT/CN2022/075153 CN2022075153W WO2022247344A1 WO 2022247344 A1 WO2022247344 A1 WO 2022247344A1 CN 2022075153 W CN2022075153 W CN 2022075153W WO 2022247344 A1 WO2022247344 A1 WO 2022247344A1
Authority
WO
WIPO (PCT)
Prior art keywords
video
feature information
sample video
sample
recognition model
Prior art date
Application number
PCT/CN2022/075153
Other languages
English (en)
French (fr)
Inventor
吴文灏
赵禹翔
Original Assignee
北京百度网讯科技有限公司
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 北京百度网讯科技有限公司 filed Critical 北京百度网讯科技有限公司
Priority to JP2022563231A priority Critical patent/JP7417759B2/ja
Priority to US17/983,208 priority patent/US20230069197A1/en
Publication of WO2022247344A1 publication Critical patent/WO2022247344A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/40—Scenes; Scene-specific elements in video content
    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/80—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level
    • G06V10/806—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level of extracted features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/40—Scenes; Scene-specific elements in video content
    • G06V20/41—Higher-level, semantic clustering, classification or understanding of video scenes, e.g. detection, labelling or Markovian modelling of sport events or news items
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/40—Scenes; Scene-specific elements in video content
    • G06V20/46—Extracting features or characteristics from the video content, e.g. video fingerprints, representative shots or key frames
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/40—Scenes; Scene-specific elements in video content
    • G06V20/49—Segmenting video sequences, i.e. computational techniques such as parsing or cutting the sequence, low-level clustering or determining units such as shots or scenes
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/70—Labelling scene content, e.g. deriving syntactic or semantic representations
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
    • H04N19/10—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding
    • H04N19/102—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding characterised by the element, parameter or selection affected or controlled by the adaptive coding
    • H04N19/132—Sampling, masking or truncation of coding units, e.g. adaptive resampling, frame skipping, frame interpolation or high-frequency transform coefficient masking

Definitions

  • the present disclosure relates to the field of artificial intelligence, specifically computer vision and deep learning technology, which can be applied to video analysis scenarios.
  • Video recognition that is, input a video and classify it according to the content of the video.
  • Video recognition is one of the most active research topics in the computer vision community. The two most important aspects for evaluating video recognition methods are classification accuracy and inference cost. Recently video recognition has achieved great success in recognition accuracy, but it is still a challenging task due to the huge computational cost.
  • the work to improve the accuracy of video recognition is mainly focused on designing a network structure that captures higher-order action semantics, and the frames input to the network are obtained by sampling at uniform or random intervals. During inference, the resulting fragment results are averaged. This method works well on short videos, but the accuracy is greatly reduced on long videos with longer and richer information.
  • Embodiments of the present disclosure provide a video recognition model training method, device, equipment, storage medium, and program product.
  • the embodiment of the present disclosure proposes a video recognition model training method, including: dividing the sample video into a plurality of sample video clips, wherein the sample video is marked with a real category label; sampling part of the sample video from the sample video clips frame, and input to the feature extraction network to obtain the feature information of the sample video segment; use the dynamic segment fusion module to perform convolution and fusion on the feature information to obtain the fusion feature information, wherein the convolution kernel of the dynamic segment fusion module varies with the video input and change; input the fusion feature information to the fully connected layer to obtain the predicted category of the sample video; adjust the parameters based on the difference between the real category label and the predicted category to obtain the video recognition model.
  • an embodiment of the present disclosure proposes a video recognition method, including: acquiring a video to be recognized; dividing the video to be recognized into multiple video segments to be recognized; sampling a portion of video frames to be recognized from the video segments to be recognized, and Input to the video recognition model to obtain the category of the video to be recognized, wherein the video recognition model is obtained by training according to the training method described in any implementation manner in the first aspect.
  • an embodiment of the present disclosure provides a video recognition model training device, including: a division module configured to divide a sample video into a plurality of sample video clips, wherein the sample video is marked with a real category label; an extraction module, It is configured to sample some sample video frames from the sample video clips, and input them to the feature extraction network to obtain the feature information of the sample video clips; the fusion module is configured to use the dynamic clip fusion module to perform convolution and fusion on the feature information to obtain fusion Feature information, wherein the convolution kernel of the dynamic segment fusion module changes with the video input; the prediction module is configured to input the fusion feature information to the fully connected layer to obtain the predicted category of the sample video; the adjustment module is configured as Based on the difference between the real category label and the predicted category, the parameters are adjusted to obtain a video recognition model.
  • an embodiment of the present disclosure provides a video recognition device, including: an acquisition module configured to acquire a video to be recognized; a division module configured to divide the video to be recognized into multiple video segments to be recognized; the recognition module , configured to sample a portion of the video frame to be recognized from the video segment to be recognized, and input it to the video recognition model to obtain the category of the video to be recognized, wherein the video recognition model is trained according to the training method described in any implementation in the first aspect get.
  • the embodiment of the present disclosure provides an electronic device, including: at least one processor; and a memory connected to the at least one processor in communication; wherein, the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by At least one processor executes, so that at least one processor can execute the method described in any implementation manner of the first aspect, or execute the method described in any implementation manner of the second aspect.
  • the embodiments of the present disclosure provide a non-transitory computer-readable storage medium storing computer instructions, the computer instructions are used to make the computer execute the method described in any implementation manner in the first aspect, or execute the method described in the second aspect. The method described in any implementation manner of the aspect.
  • the embodiments of the present disclosure provide a computer program product, including a computer program.
  • the computer program When executed by a processor, it implements the method described in any implementation manner in the first aspect, or implements the method described in any implementation manner in the second aspect. A method described in an implementation mode.
  • the video recognition model training method, device, device, storage medium, and program product provided by the embodiments of the present disclosure enable the convolution kernel of the video recognition model to change according to different video inputs during training and reasoning by designing a dynamic segment fusion module. changes, thereby improving the recognition accuracy.
  • the video recognition model adopts the recognition method of dynamic convolution fusion.
  • the convolution kernel parameters of the fusion segment can change with the input video, realizing more accurate time domain perception than simply using a convolution kernel, and improving the recognition accuracy. At the same time, it will not increase the computational complexity. In particular, it can improve the recognition accuracy of long videos with longer and richer information. It can be used for classification of medium and long videos, content classification of film and television dramas, etc.
  • Fig. 1 is a flowchart of an embodiment of a video recognition model training method according to the present disclosure
  • FIG. 2 is a flow chart of another embodiment of a video recognition model training method according to the present disclosure.
  • Fig. 3 is a scene diagram that can realize the video recognition model training method of the embodiment of the present disclosure
  • Fig. 4 is the structural representation of video recognition model
  • FIG. 5 is a schematic diagram of the structure of the DSA Block
  • FIG. 6 is a flowchart of an embodiment of a video recognition method according to the present disclosure.
  • FIG. 7 is a schematic structural diagram of an embodiment of a video recognition model training device according to the present disclosure.
  • Fig. 8 is a schematic structural diagram of an embodiment of a video recognition device according to the present disclosure.
  • FIG. 9 is a block diagram of an electronic device used to implement the video recognition model training method or the video recognition method according to an embodiment of the present disclosure.
  • FIG. 1 shows a flow 100 of an embodiment of a method for training a video recognition model according to the present disclosure.
  • the video recognition model training method comprises the following steps:
  • Step 101 divide the sample video into a plurality of sample video segments.
  • the execution subject of the video recognition model training method may obtain a sample video set.
  • the execution subject may divide the sample video into a plurality of sample video clips.
  • the sample video set may include a large number of sample videos labeled with real category labels.
  • the true category label of the sample video annotation can be obtained by using other video recognition models, or by manual classification, which is not limited here.
  • the sample video can be divided into sample video segments in various ways.
  • the sample video is evenly divided according to the video length to obtain multiple sample video clips of the same length.
  • the sample video is divided according to a fixed length to obtain multiple sample video clips with a fixed length.
  • the sample video is randomly divided to obtain a plurality of sample video clips with random lengths.
  • Step 102 sampling part of the sample video frames from the sample video clips, and inputting them to the feature extraction network to obtain the feature information of the sample video clips.
  • the execution subject may sample some sample video frames from the sample video clips and input them to the feature extraction network to obtain feature information of the sample video clips. Only a part of the sample video frames is sampled and input to the feature extraction network for feature extraction, which can reduce the training workload and shorten the training time.
  • the feature extraction network can be used to extract features from videos, including but not limited to various neural networks for feature extraction.
  • CNN Convolutional Neural Network, Convolutional Neural Network
  • the sample video segment can sample the sample video frame in various ways. For example, uniformly spaced sampling is performed on the sample video segment to obtain a plurality of uniformly spaced sample video frames. For another example, random sampling is performed on the sample video segment to obtain a plurality of sample video frames at random intervals.
  • Step 103 using the dynamic segment fusion module to perform convolution fusion on the feature information to obtain the fusion feature information.
  • the above-mentioned execution subject may use a dynamic segment fusion module (DSA Module) to perform convolution fusion on feature information to obtain fusion feature information.
  • DSA Module dynamic segment fusion module
  • the convolution kernel of the dynamic segment fusion module can vary with different video inputs.
  • the dynamic segment fusion module For the differences in feature information of different videos, especially the differences in feature channels, the dynamic segment fusion module generates a dynamic convolution kernel.
  • the convolution kernel can change with the input video and is related to the input channel.
  • the convolution kernel is convoluted and fused with the feature information of each video segment of the video, so as to realize the perception and modeling of the long-term video domain.
  • a video recognition model may include multiple residual layers, and a dynamic segment fusion module may be set inside the residual layer.
  • a dynamic segment fusion module may be set inside the residual layer.
  • at least one dynamic segment fusion module may be arranged at intervals in multiple residual layers of the video recognition model.
  • video recognition models include Res2, Res3, Res4, and Res5. Two dynamic fusion modules are respectively set inside Res3 and Res5.
  • Step 104 input the fused feature information to the fully connected layer to obtain the predicted category of the sample video.
  • the above-mentioned executor can input the fusion feature information into the fully connected layer for classification, and then the predicted category of the sample video can be obtained.
  • the fully connected layer can output the scores of the sample videos belonging to each preset category.
  • Step 105 parameter adjustment is performed based on the difference between the real category label and the predicted category to obtain a video recognition model.
  • the above execution subject may perform parameter adjustment based on the difference between the real category label and the predicted category to obtain a video recognition model.
  • the purpose of adjusting the parameters is to make the difference between the real category label and the predicted category small enough.
  • the execution subject can first calculate the cross-entropy loss based on the real category label and the predicted category; then use SGD (Stochastic Gradient Descent, stochastic gradient descent) to optimize the cross-entropy loss, and continuously Parameters are updated until the cross-entropy loss converges to obtain a video recognition model.
  • SGD Spochastic Gradient Descent, stochastic gradient descent
  • the convolution kernel of the video recognition model can change with different video inputs during training and reasoning, thereby improving the recognition accuracy.
  • the video recognition model adopts the recognition method of dynamic convolution fusion.
  • the convolution kernel parameters of the fusion segment can change with the input video, realizing more accurate time domain perception than simply using a convolution kernel, and improving the recognition accuracy.
  • it will not increase the computational complexity.
  • it can improve the recognition accuracy of long videos with longer and richer information. It can be applied to the classification of medium and long videos, the content classification of film and television dramas, etc.
  • FIG. 2 shows a flow 200 of another embodiment of the video recognition model training method according to the present disclosure.
  • the video recognition model training method comprises the following steps:
  • step 201 the sample video is evenly divided according to the length of the video to obtain a plurality of sample video clips.
  • the execution subject of the video recognition model training method may obtain a sample video set.
  • the execution subject may evenly divide the sample videos according to the length of the videos to obtain multiple sample video clips. For example, for a sample video of 10 seconds, it is evenly divided every 2 seconds to obtain five sample video clips of 2 seconds.
  • the sample video set may include a large number of sample videos labeled with real category labels.
  • the true category label of the sample video annotation can be obtained by using other video recognition models, or by manual classification, which is not limited here.
  • step 202 the sample video clips are sampled at even intervals to obtain some sample video frames, which are input to the feature extraction network to obtain feature information of the sample video clips.
  • the execution subject can sample the sample video clips at even intervals to obtain some sample video frames, and input them to the feature extraction network to obtain the features of the sample video clips. information. Only a part of the sample video frames is sampled and input to the feature extraction network for feature extraction, which can reduce the training workload and shorten the training time. For example, for a sample video clip of 2 seconds, uniform sampling is performed every 0.25 seconds to obtain 8 sample video frames.
  • the feature extraction network can be used to extract features from videos, including but not limited to various neural networks for feature extraction. For example, CNN.
  • the sample video is evenly divided according to the length of the video, and then the divided sample video segments are sampled at even intervals, so that the feature extraction network can extract the feature information of each position of the sample video.
  • step 203 the feature information is divided into first feature information and second feature information in the channel dimension.
  • the execution subject may divide feature information into first feature information and second feature information in a channel dimension.
  • the first feature information and the second feature information correspond to different channel dimensions.
  • the execution subject may divide feature information into first feature information and second feature information in a channel dimension according to a preset hyperparameter ⁇ .
  • the channel dimension of the first feature information may be ⁇ C
  • the channel dimension of the second feature information may be (1- ⁇ )C
  • C is the channel dimension of feature information.
  • ⁇ is a hyperparameter, and its value range is (0,1). Since the first feature information needs to be convoluted, the second feature information only needs to be concatenated. Therefore, by adjusting the hyperparameter ⁇ , the amount of convolution calculation can be controlled.
  • the value range of the hyperparameter ⁇ is set to (0,0.5), which can reduce the amount of convolution calculations.
  • Step 204 using the convolution kernel generation branch network to determine the convolution kernel corresponding to the sample video.
  • the execution subject may use the convolution kernel generation branch network to determine the convolution kernel corresponding to the sample video.
  • the dynamic segment fusion module may include a convolution kernel generation branch network.
  • the convolution kernel generation branch network can be used to generate convolution kernels.
  • the convolution kernel can be changed as the input video changes.
  • the execution subject may first calculate the channel dimension ⁇ C of the first characteristic information, the segment number U of the sample video, the number T of sampling frames of the sample video segment, and the height H of the sample video frame and the product of width W ⁇ C ⁇ U ⁇ T ⁇ H ⁇ W; then the product ⁇ C ⁇ U ⁇ T ⁇ H ⁇ W is input to the convolution kernel generation branch network, so as to quickly obtain the convolution kernel corresponding to the sample video.
  • the convolution kernel generation branch network can include a GAP (Global Average Pooling, global average pooling layer) and two FC (Full Connected Layer, fully connected layer).
  • Step 205 using the convolution kernel corresponding to the sample video to convolve the first feature information to obtain a convolution result.
  • the execution subject may use the convolution kernel corresponding to the sample video to perform convolution on the first feature information to obtain a convolution result.
  • Step 206 concatenate the convolution result with the second feature information to obtain the fusion feature.
  • the execution subject may concatenate the convolution result with the second feature information to obtain the fused feature.
  • the feature information is divided into the first feature information and the second feature information in the channel dimension, and only the first feature information is convolved and spliced with the second feature information to obtain the fusion feature, which can reduce the amount of convolution calculation .
  • Step 207 input the fused feature information to the fully connected layer to obtain the predicted category of the sample video.
  • step 208 parameters are adjusted based on the difference between the real category label and the predicted category to obtain a video recognition model.
  • steps 207-208 have been introduced in detail in steps 104-105 in the embodiment shown in FIG. 1 , and will not be repeated here.
  • the video recognition model training method in this embodiment highlights the steps of video division, video frame sampling and convolution fusion. Therefore, the solution described in this embodiment divides the sample video evenly according to the length of the video, and then samples the divided sample video segments at even intervals, so that the feature extraction network can extract the feature information of each position of the sample video.
  • the feature information is divided into the first feature information and the second feature information in the channel dimension, the first feature information is convolved, and spliced with the second feature information to obtain the fusion feature, which can reduce the amount of convolution calculation.
  • Fig. 3 shows the scene diagram that can realize the video recognition model training method of the embodiment of the present disclosure.
  • the sample video is evenly divided into 4 sample video segments (Snippets), and 4 video frames are sampled at even intervals from each sample video segment.
  • the 4 video frames of the 4 sample video clips are respectively input to CNN Layers to obtain the feature information of the 4 sample video clips.
  • the DSA Module is used to perform convolution fusion on the feature information of the four sample video clips, and the obtained fusion features are continuously input to the CNN Layers for processing.
  • FIG. 4 it shows a schematic structural diagram of a video recognition model.
  • the video recognition model may include a convolutional layer, multiple residual layers, and a fully connected layer, and multiple residual layers may be provided with dynamic segment fusion modules at intervals.
  • the video recognition model includes Conv1, Res2, Res3, Res4, Res5 and FC.
  • the Snippets of the sample video are processed by Conv1, Res2, Res3, Res4, Res5, and FC to obtain the predicted category of the sample video (score belonging to each preset category).
  • Two dynamic fusion modules are respectively set inside Res3 and Res5.
  • Figure 4 only shows the structure of Res3, including 2 Res Blocks and 2 DSA Blocks.
  • the structure of Res5 is the same as that of Res3, which is not shown in FIG. 3 .
  • FIG. 5 shows a schematic diagram of the structure of the DSA Block.
  • Figure 5 shows two DSA Blocks.
  • (a) in Figure 5 shows the DSA Block (for TSM), which is a 2D DSA Block.
  • (b) in Figure 5 shows DSA Block (for I3D), which is a 3D DSA Block.
  • (c) in Figure 5 shows a schematic structural diagram of the DSA Module in DSA Block (for TSM) and DSA Block (for I3D).
  • DSA Module includes a GAP and two FCs.
  • Feature information is divided into first feature information ⁇ C and second feature information (1- ⁇ )C in the channel dimension.
  • the product ⁇ C ⁇ U ⁇ T ⁇ H ⁇ W is input to GAP to obtain ⁇ C ⁇ U.
  • ⁇ C ⁇ U is input to FC to obtain ⁇ C ⁇ aU.
  • ⁇ C ⁇ aU is input to FC to obtain ⁇ C ⁇ L.
  • ⁇ C ⁇ L is convolved with ⁇ C ⁇ U ⁇ T ⁇ H ⁇ W and concatenated with (1- ⁇ )C ⁇ U ⁇ T ⁇ H ⁇ W.
  • FIG. 6 it shows a flow 600 of an embodiment of the video recognition method according to the present disclosure.
  • the video recognition method includes the following steps:
  • Step 601 acquire the video to be recognized.
  • the executing subject of the video recognition method may acquire the video to be recognized.
  • Step 602 divide the video to be recognized into multiple video segments to be recognized.
  • the execution subject may divide the video to be recognized into multiple video segments to be recognized.
  • the division granularity of the video to be recognized is greater than the division granularity of the sample video used for training the video recognition model.
  • the number of sample videos used to train the video recognition model is huge, and reducing the granularity of the sample videos can shorten the training time.
  • increasing the division granularity of the video to be recognized can improve the recognition accuracy. For example, for a sample video of 10 seconds, it is evenly divided every 2 seconds to obtain five sample video clips of 2 seconds. For the 10-second video to be recognized, it is evenly divided every 1 second to obtain 10 1-second video segments to be recognized.
  • Step 603 Sampling a part of the video frame to be recognized from the video segment to be recognized, and inputting it into the video recognition model to obtain the category of the video to be recognized.
  • some video frames to be recognized are sampled from video segments to be recognized, input to the video recognition model for prediction, and the prediction results are aggregated to obtain the category of the video to be recognized.
  • sampling method of the video segment to be identified refer to the sampling method of the sample video segment, which will not be repeated here.
  • the video recognition model can be used for video classification, which is obtained through training using the training method provided by any of the implementations in Fig. 1-Fig. 2, and will not be repeated here.
  • the video recognition method provided by the embodiment of the present disclosure provides an efficient video recognition method based on dynamic segment fusion.
  • the convolution kernel of the video recognition model can be trained and reasoned with the video input different, thus improving the recognition accuracy.
  • the video recognition model adopts the recognition method of dynamic convolution fusion.
  • the convolution kernel parameters of the fusion segment can change with the input video, realizing more accurate time domain perception than simply using a convolution kernel, and improving the recognition accuracy. At the same time, it will not increase the computational complexity. In particular, it can improve the recognition accuracy of long videos with longer and richer information. It can be used for classification of medium and long videos, content classification of film and television dramas, etc.
  • the present disclosure provides an embodiment of a video recognition model training device, which corresponds to the method embodiment shown in FIG. 1 , and the device Specifically, it can be applied to various electronic devices.
  • the video recognition model training apparatus 700 of this embodiment may include: a division module 701 , an extraction module 702 , a fusion module 703 , a prediction module 704 and an adjustment module 705 .
  • the division module 701 is configured to divide the sample video into a plurality of sample video clips, wherein the sample video is marked with a real category label
  • the extraction module 702 is configured to sample some sample video frames from the sample video clips, and input To the feature extraction network to obtain the feature information of the sample video segment
  • the fusion module 703 is configured to use the dynamic segment fusion module to perform convolution and fusion of the feature information to obtain the fusion feature information, wherein the convolution kernel of the dynamic segment fusion module follows the video
  • the prediction module 704 is configured to input the fusion feature information to the fully connected layer to obtain the predicted category of the sample video
  • the adjustment module 705 is configured to perform parameter adjustment based on the difference between the real category label and the predicted category, Get the video recognition model.
  • the specific processing of the division module 701, the extraction module 702, the fusion module 703, the prediction module 704 and the adjustment module 705 and the technical effects brought by them can refer to FIG. 1 respectively. Relevant descriptions of steps 101-105 in the embodiment will not be repeated here.
  • the fusion module 703 includes: a division submodule configured to divide feature information into first feature information and second feature information in the channel dimension; a determination submodule configured The method is to use the convolution kernel to generate a branch network to determine the convolution kernel corresponding to the sample video; the convolution sub-module is configured to use the convolution kernel corresponding to the sample video to convolve the first feature information to obtain a convolution result; the splicing sub-module, It is configured to concatenate the convolution result with the second feature information to obtain fusion features.
  • the division submodule is further configured to: divide the feature information into first feature information and second feature information in the channel dimension according to the preset hyperparameter ⁇ , wherein the first The channel dimension of the feature information is ⁇ C, the channel dimension of the second feature information is (1- ⁇ )C, and C is the channel dimension of the feature information.
  • the determining submodule is further configured to: calculate the channel dimension ⁇ C of the first feature information, the number of segments of the sample video, the number of sampling frames of the sample video segment, and the height of the sample video frame. and the product of the width; input the product to the convolution kernel generation branch network to obtain the convolution kernel corresponding to the sample video.
  • the convolution kernel generation branch network includes a global average pooling layer and two fully connected layers.
  • the video recognition model includes multiple residual layers, and at least one dynamic segment fusion module is arranged at intervals between the multiple residual layers.
  • the division module 701 is further configured to: uniformly divide the sample video according to the length of the video to obtain a plurality of sample video clips; and the extraction module 1002 is further configured to: divide the sample video Segments are evenly spaced to obtain partial sample video frames.
  • the adjustment module 705 is further configured to: calculate the cross-entropy loss based on the real category label and the predicted category; use stochastic gradient descent to optimize the cross-entropy loss, and continuously update the parameters until The cross-entropy loss converges, resulting in a video recognition model.
  • the present disclosure provides an embodiment of a video recognition device, which corresponds to the method embodiment shown in FIG. 6 , and the device can specifically Used in various electronic equipment.
  • the video recognition apparatus 800 of this embodiment may include: an acquisition module 801 , a division module 802 and a recognition module 803 .
  • the acquisition module 801 is configured to acquire the video to be identified;
  • the division module 802 is configured to divide the video to be identified into a plurality of video segments to be identified;
  • the identification module 803 is configured to sample a portion of the video to be identified from the video segments to be identified Recognize the video frame and input it into the video recognition model to obtain the category of the video to be recognized, wherein the video recognition model is obtained by training according to the training method described in any one of the embodiments in FIGS. 1-2 .
  • the specific processing of the acquisition module 801, the division module 802 and the recognition module 803 and the technical effects brought by them can refer to the relevant steps 601-603 in the corresponding embodiment in FIG. description and will not be repeated here.
  • the division granularity of the video to be recognized is greater than the division granularity of the sample video used for training the video recognition model.
  • the acquisition, storage and application of the user's personal information involved are in compliance with relevant laws and regulations, and do not violate public order and good customs.
  • the present disclosure also provides an electronic device, a readable storage medium, and a computer program product.
  • FIG. 9 shows a schematic block diagram of an example electronic device 900 that may be used to implement embodiments of the present disclosure.
  • Electronic device is intended to represent various forms of digital computers, such as laptops, desktops, workstations, personal digital assistants, servers, blade servers, mainframes, and other suitable computers.
  • Electronic devices may also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices, and other similar computing devices.
  • the components shown herein, their connections and relationships, and their functions, are by way of example only, and are not intended to limit implementations of the disclosure described and/or claimed herein.
  • the device 900 includes a computing unit 901 that can execute according to a computer program stored in a read-only memory (ROM) 902 or loaded from a storage unit 908 into a random-access memory (RAM) 903. Various appropriate actions and treatments. In the RAM 903, various programs and data necessary for the operation of the device 900 can also be stored.
  • the computing unit 901, ROM 902, and RAM 903 are connected to each other through a bus 904.
  • An input/output (I/O) interface 905 is also connected to the bus 904 .
  • the I/O interface 905 includes: an input unit 906, such as a keyboard, a mouse, etc.; an output unit 907, such as various types of displays, speakers, etc.; a storage unit 908, such as a magnetic disk, an optical disk, etc. ; and a communication unit 909, such as a network card, a modem, a wireless communication transceiver, and the like.
  • the communication unit 909 allows the device 900 to exchange information/data with other devices over a computer network such as the Internet and/or various telecommunication networks.
  • the computing unit 901 may be various general-purpose and/or special-purpose processing components having processing and computing capabilities. Some examples of computing units 901 include, but are not limited to, central processing units (CPUs), graphics processing units (GPUs), various dedicated artificial intelligence (AI) computing chips, various computing units that run machine learning model algorithms, digital signal processing processor (DSP), and any suitable processor, controller, microcontroller, etc.
  • the calculation unit 901 executes various methods and processes described above, such as a video recognition model training method.
  • the video recognition model training method may be implemented as a computer software program tangibly embodied on a machine-readable medium, such as storage unit 908 .
  • part or all of the computer program may be loaded and/or installed on the device 900 via the ROM 902 and/or the communication unit 909.
  • the computer program When the computer program is loaded into the RAM 903 and executed by the computing unit 901, one or more steps of the video recognition model training method described above can be performed.
  • the computing unit 901 may be configured in any other appropriate way (for example, by means of firmware) to execute the video recognition model training method.
  • Various implementations of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips Implemented in a system of systems (SOC), load programmable logic device (CPLD), computer hardware, firmware, software, and/or combinations thereof.
  • FPGAs field programmable gate arrays
  • ASICs application specific integrated circuits
  • ASSPs application specific standard products
  • SOC system of systems
  • CPLD load programmable logic device
  • computer hardware firmware, software, and/or combinations thereof.
  • programmable processor can be special-purpose or general-purpose programmable processor, can receive data and instruction from storage system, at least one input device, and at least one output device, and transmit data and instruction to this storage system, this at least one input device, and this at least one output device an output device.
  • Program codes for implementing the methods of the present disclosure may be written in any combination of one or more programming languages. These program codes may be provided to a processor or controller of a general-purpose computer, a special purpose computer, or other programmable data processing devices, so that the program codes, when executed by the processor or controller, make the functions/functions specified in the flow diagrams and/or block diagrams Action is implemented.
  • the program code may execute entirely on the machine, partly on the machine, as a stand-alone software package partly on the machine and partly on a remote machine or entirely on the remote machine or server.
  • a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device.
  • a machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium.
  • a machine-readable medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing.
  • machine-readable storage media would include one or more wire-based electrical connections, portable computer discs, hard drives, random access memory (RAM), read only memory (ROM), erasable programmable read only memory (EPROM or flash memory), optical fiber, compact disk read only memory (CD-ROM), optical storage, magnetic storage, or any suitable combination of the foregoing.
  • RAM random access memory
  • ROM read only memory
  • EPROM or flash memory erasable programmable read only memory
  • CD-ROM compact disk read only memory
  • magnetic storage or any suitable combination of the foregoing.
  • the systems and techniques described herein can be implemented on a computer having a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user. ); and a keyboard and pointing device (eg, a mouse or a trackball) through which a user can provide input to the computer.
  • a display device e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor
  • a keyboard and pointing device eg, a mouse or a trackball
  • Other kinds of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and can be in any form (including Acoustic input, speech input or, tactile input) to receive input from the user.
  • the systems and techniques described herein can be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., as a a user computer having a graphical user interface or web browser through which a user can interact with embodiments of the systems and techniques described herein), or including such backend components, middleware components, Or any combination of front-end components in a computing system.
  • the components of the system can be interconnected by any form or medium of digital data communication, eg, a communication network. Examples of communication networks include: Local Area Network (LAN), Wide Area Network (WAN) and the Internet.
  • a computer system may include clients and servers.
  • Clients and servers are generally remote from each other and typically interact through a communication network.
  • the relationship of client and server arises by computer programs running on the respective computers and having a client-server relationship to each other.
  • the server can be a cloud server, a server of a distributed system, or a server combined with a blockchain.
  • steps may be reordered, added or deleted using the various forms of flow shown above.
  • each step described in the present disclosure may be executed in parallel, sequentially, or in a different order, as long as the desired result of the technical solution provided by the present disclosure can be achieved, no limitation is imposed herein.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Multimedia (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Databases & Information Systems (AREA)
  • Medical Informatics (AREA)
  • Computational Linguistics (AREA)
  • Signal Processing (AREA)
  • General Engineering & Computer Science (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • Biophysics (AREA)
  • Molecular Biology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Data Mining & Analysis (AREA)
  • Image Analysis (AREA)

Abstract

本公开提供了一种视频识别模型训练方法、装置、设备、存储介质以及程序产品，涉及人工智能领域，具体为计算机视觉和深度学习技术，可应用于视频分析场景下。该方法的一具体实施方式包括：将样本视频划分成多个样本视频片段；从样本视频片段中采样部分样本视频帧，并输入至特征提取网络，得到样本视频片段的特征信息；利用动态片段融合模块对特征信息进行卷积融合，得到融合特征信息，其中，动态片段融合模块的卷积核随视频输入的不同而变化；将融合特征信息输入至全连接层，得到样本视频的预测类别；基于真实类别标签与预测类别的差异进行参数调整，得到视频识别模型。该实施方式提高了视频识别模型的识别精度。

Description

视频识别模型训练方法、装置、设备以及存储介质
本专利申请要求于2021年5月28日提交的、申请号为202110589375.6、发明名称为“视频识别模型训练方法、装置、设备以及存储介质”的中国专利申请的优先权，该申请的全文以引用的方式并入本申请中。
技术领域
本公开涉及人工智能领域，具体为计算机视觉和深度学习技术，可应用于视频分析场景下。
背景技术
视频识别，即输入一段视频根据视频内容对其进行分类。视频识别是计算机视觉社区中最活跃的研究主题之一。评估视频识别方法的两个最重要方面是分类准确性和推理成本。最近视频识别已经在识别精度上取得了巨大的成功，但是由于庞大的计算成本，它仍然是一项具有挑战性的任务。
目前针对深度学习相关的方法，提高视频识别精度的工作主要集中在设计捕获更高阶动作语意的网络结构上，而输入网络的帧通过均匀或随机间隔采样获得。在推理过程中，对得到的片段结果进行平均。该方法在短视频上效果尚可，但是在拥有更长更丰富信息的长视频上精度则大打折扣。
发明内容
本公开实施例提出了一种视频识别模型训练方法、装置、设备、存储介质以及程序产品。
第一方面，本公开实施例提出了一种视频识别模型训练方法，包括：将样本视频划分成多个样本视频片段，其中，样本视频标注有真实类别标签；从样本视频片段中采样部分样本视频帧，并输入至特征提取网络， 得到样本视频片段的特征信息；利用动态片段融合模块对特征信息进行卷积融合，得到融合特征信息，其中，动态片段融合模块的卷积核随视频输入的不同而变化；将融合特征信息输入至全连接层，得到样本视频的预测类别；基于真实类别标签与预测类别的差异进行参数调整，得到视频识别模型。
第二方面，本公开实施例提出了一种视频识别方法，包括：获取待识别视频；将待识别视频划分成多个待识别视频片段；从待识别视频片段中采样部分待识别视频帧，并输入至视频识别模型，得到待识别视频的类别，其中，视频识别模型根据第一方面中任一实现方式描述的训练方法训练获得。
第三方面，本公开实施例提出了一种视频识别模型训练装置，包括：划分模块，被配置成将样本视频划分成多个样本视频片段，其中，样本视频标注有真实类别标签；提取模块，被配置成从样本视频片段中采样部分样本视频帧，并输入至特征提取网络，得到样本视频片段的特征信息；融合模块，被配置成利用动态片段融合模块对特征信息进行卷积融合，得到融合特征信息，其中，动态片段融合模块的卷积核随视频输入的不同而变化；预测模块，被配置成将融合特征信息输入至全连接层，得到样本视频的预测类别；调整模块，被配置成基于真实类别标签与预测类别的差异进行参数调整，得到视频识别模型。
第四方面，本公开实施例提出了一种视频识别装置，包括：获取模块，被配置成获取待识别视频；划分模块，被配置成将待识别视频划分成多个待识别视频片段；识别模块，被配置成从待识别视频片段中采样部分待识别视频帧，并输入至视频识别模型，得到待识别视频的类别，其中，视频识别模型根据第一方面中任一实现方式描述的训练方法训练获得。
第五方面，本公开实施例提出了一种电子设备，包括：至少一个处理器；以及与至少一个处理器通信连接的存储器；其中，存储器存储有可被至少一个处理器执行的指令，指令被至少一个处理器执行，以使至少一个处理器能够执行如第一方面中任一实现方式描述的方法，或者执行如第二方面中任一实现方式描述的方法。
第六方面，本公开实施例提出了一种存储有计算机指令的非瞬时 计算机可读存储介质，计算机指令用于使计算机执行如第一方面中任一实现方式描述的方法，或者执行如第二方面中任一实现方式描述的方法。
第七方面，本公开实施例提出了一种计算机程序产品，包括计算机程序，计算机程序在被处理器执行时实现如第一方面中任一实现方式描述的方法，或者实现如第二方面中任一实现方式描述的方法。
本公开实施例提供的视频识别模型训练方法、装置、设备、存储介质以及程序产品，通过设计动态片段融合模块，使视频识别模型的卷积核能够在训练和推理中随着视频输入的不同而变化，从而提高了识别精度。视频识别模型采用动态卷积融合的识别方式，融合片段的卷积核参数能够随着输入的视频而变化，实现比单纯使用一种卷积核更为精准的时域感知，在提升识别精度的同时不会增加计算复杂度。尤其能够提升拥有更长更丰富的信息的长视频的识别精度。可用于中长视频分类、影视剧内容分类等。
应当理解，本部分所描述的内容并非旨在标识本公开的实施例的关键或重要特征，也不用于限制本公开的范围。本公开的其它特征将通过以下的说明书而变得容易理解。
附图说明
通过阅读参照以下附图所作的对非限制性实施例所作的详细描述，本公开的其它特征、目的和优点将会变得更明显。附图用于更好地理解本方案，不构成对本公开的限定。其中：
图1是根据本公开的视频识别模型训练方法的一个实施例的流程图；
图2是根据本公开的视频识别模型训练方法的又一个实施例的流程图；
图3是可以实现本公开实施例的视频识别模型训练方法的场景图；
图4是视频识别模型的结构示意图；
图5是DSA Block的结构示意图；
图6是根据本公开的视频识别方法的一个实施例的流程图；
图7是根据本公开的视频识别模型训练装置的一个实施例的结构示意图；
图8是根据本公开的视频识别装置的一个实施例的结构示意图；
图9是用来实现本公开实施例的视频识别模型训练方法或视频识别方法的电子设备的框图。
具体实施方式
以下结合附图对本公开的示范性实施例做出说明，其中包括本公开实施例的各种细节以助于理解，应当将它们认为仅仅是示范性的。因此，本领域普通技术人员应当认识到，可以对这里描述的实施例做出各种改变和修改，而不会背离本公开的范围和精神。同样，为了清楚和简明，以下的描述中省略了对公知功能和结构的描述。
需要说明的是，在不冲突的情况下，本公开中的实施例及实施例中的特征可以相互组合。下面将参考附图并结合实施例来详细说明本公开。
图1示出了根据本公开的视频识别模型训练方法的一个实施例的流程100。该视频识别模型训练方法包括以下步骤：
步骤101，将样本视频划分成多个样本视频片段。
在本实施例中，视频识别模型训练方法的执行主体可以获取样本视频集。对于样本视频集中的样本视频，上述执行主体可以将样本视频划分成多个样本视频片段。
其中，样本视频集可以包括大量标注有真实类别标签的样本视频。样本视频标注的真实类别标签可以利用其他视频识别模型分类得到，也可以人工分类得到，这里不进行限定。
这里，样本视频可以通过多种方式来划分样本视频片段。例如，根据视频长度对样本视频进行均匀划分，得到多个相同长度的样本视频片段。又例如，按照固定长度对样本视频进行划分，得到多个固定长度的样本视频片段。再例如，对样本视频进行随机划分，得到多个随机长度的样本视频片段。
步骤102，从样本视频片段中采样部分样本视频帧，并输入至特征提 取网络，得到样本视频片段的特征信息。
在本实施例中，对于多个样本视频片段中的样本视频片段，上述执行主体可以从样本视频片段中采样部分样本视频帧，并输入至特征提取网络，得到样本视频片段的特征信息。仅采样部分样本视频帧，输入特征提取网络进行特征提取，可以降低训练工作量，缩短训练时间。
其中，特征提取网络可以用于从视频中提取特征，包括但不限于各种用于提取特征的神经网络。例如，CNN(Convolutional Neural Network，卷积神经网络)。
这里，样本视频片段可以通过多种方式来采样样本视频帧。例如，对样本视频片段进行均匀间隔采样，得到多个均匀间隔的样本视频帧。又例如，对样本视频片段进行随机采样，得到多个随机间隔的样本视频帧。
步骤103，利用动态片段融合模块对特征信息进行卷积融合，得到融合特征信息。
在本实施例中，上述执行主体可以利用动态片段融合模块(DSA Module)对特征信息进行卷积融合，得到融合特征信息。
其中，动态片段融合模块的卷积核可以随视频输入的不同而变化。对于不同视频在特征信息上表现出的差异，尤其是在特征通道上表现出的差异，动态片段融合模块生成一种动态的卷积核。卷积核可以随输入视频的变化而改变，且与输入通道相关。卷积核对与视频的各个视频片段的特征信息进行卷积融合，从而实现对视频长时域的感知和建模。
通常，视频识别模型可以包括多个残差层，残差层内部可以设置动态片段融合模块。实践中，设置的动态融合模块越多，融合次数越多，识别精度越高，然而计算量越大。因此，可以综合考虑识别精度需求和计算量需求，确定设置的动态融合模块的数目。可选地，视频识别模型的多个残差层可以间隔设置至少一个动态片段融合模块。例如，视频识别模型包括Res2、Res3、Res4和Res5。在Res3和Res5内部分别设置两个动态融合模块。
步骤104，将融合特征信息输入至全连接层，得到样本视频的预测类别。
在本实施例中，上述执行主体可以将融合特征信息输入至全连接层 进行分类，即可得到样本视频的预测类别。其中，全连接层可以输出样本视频属于各个预设类别的得分。
步骤105，基于真实类别标签与预测类别的差异进行参数调整，得到视频识别模型。
在本实施例中，上述执行主体可以基于真实类别标签与预测类别的差异进行参数调整，得到视频识别模型。其中，调整参数的目的是为了使真实类别标签与预测类别的差异足够小。
在本实施例的一些可选的实现方式中，上述执行主体可以首先基于真实类别标签和预测类别计算交叉熵损失；然后使用SGD(Stochastic Gradient Descent，随机梯度下降)对交叉熵损失进行优化，不断进行参数更新，直至交叉熵损失收敛，得到视频识别模型。
本公开实施例提供的视频识别模型训练方法，通过设计动态片段融合模块，使视频识别模型的卷积核能够在训练和推理中随着视频输入的不同而变化，从而提高了识别精度。视频识别模型采用动态卷积融合的识别方式，融合片段的卷积核参数能够随着输入的视频而变化，实现比单纯使用一种卷积核更为精准的时域感知，在提升识别精度的同时不会增加计算复杂度。尤其能够提升拥有更长更丰富的信息的长视频的识别精度。可应用于中长视频分类、影视剧内容分类等方面。
继续参考图2，其示出了根据本公开的视频识别模型训练方法的又一个实施例的流程200。该视频识别模型训练方法包括以下步骤：
步骤201，根据视频长度对样本视频进行均匀划分，得到多个样本视频片段。
在本实施例中，视频识别模型训练方法的执行主体可以获取样本视频集。对于样本视频集中的样本视频，上述执行主体可以根据视频长度对样本视频进行均匀划分，得到多个样本视频片段。例如，对于10秒的样本视频，每隔2秒进行均匀划分，得到5个2秒的样本视频片段。
其中，样本视频集可以包括大量标注有真实类别标签的样本视频。样本视频标注的真实类别标签可以利用其他视频识别模型分类得到，也可以人工分类得到，这里不进行限定。
步骤202，对样本视频片段进行均匀间隔采样，得到部分样本视频帧， 并输入至特征提取网络，得到样本视频片段的特征信息。
在本实施例中，对于多个样本视频片段中的样本视频片段，上述执行主体可以对样本视频片段进行均匀间隔采样，得到部分样本视频帧，并输入至特征提取网络，得到样本视频片段的特征信息。仅采样部分样本视频帧，输入特征提取网络进行特征提取，可以降低训练工作量，缩短训练时间。例如，对于2秒的样本视频片段，每隔0.25秒进行均匀采样，得到8帧样本视频帧。
其中，特征提取网络可以用于从视频中提取特征，包括但不限于各种用于提取特征的神经网络。例如，CNN。
这里，先根据视频长度对样本视频均匀划分，再对划分得到的样本视频片段均匀间隔采样，从而使得特征提取网络可以提取到样本视频的各个位置的特征信息。
步骤203，将特征信息在通道维度上划分成第一特征信息和第二特征信息。
在本实施例中，上述执行主体可以将特征信息在通道维度上划分成第一特征信息和第二特征信息。其中，第一特征信息和第二特征信息对应不同的通道维度。
在本实施例的一些可选的实现方式中，上述执行主体可以按照预设超参数β将特征信息在通道维度上划分成第一特征信息和第二特征信息。其中，第一特征信息的通道维度可以是βC，第二特征信息的通道维度可以是(1-β)C。C是特征信息的通道维度。β是超参数，其取值范围是(0,1)。由于第一特征信息需要进行卷积操作，第二特征信息仅需要进行拼接操作，因此，通过调整超参数β，可以控制卷积计算量。通常，超参数β的取值范围设置为(0,0.5)，可以降低卷积计算量。
步骤204，利用卷积核生成分支网络确定样本视频对应的卷积核。
在本实施例中，上述执行主体可以利用卷积核生成分支网络确定样本视频对应的卷积核。
其中，动态片段融合模块(DSA Module)可以包括卷积核生成分支网络。卷积核生成分支网络可以用于生成卷积核。卷积核可以随输入视频的变化而改变。
在本实施例的一些可选的实现方式中，上述执行主体可以首先计算第一特征信息的通道维度βC、样本视频的片段数U、样本视频片段的采样帧数T、样本视频帧的高H和宽W的乘积βC×U×T×H×W；然后将乘积βC×U×T×H×W输入至卷积核生成分支网络，从而快速得到样本视频对应的卷积核。其中，卷积核生成分支网络可以包括一个GAP(Global Average Pooling，全局平均池化层)和两个FC(Full Connected Layer，全连接层)。
步骤205，利用样本视频对应的卷积核对第一特征信息进行卷积，得到卷积结果。
在本实施例中，上述执行主体可以利用样本视频对应的卷积核对第一特征信息进行卷积，得到卷积结果。
步骤206，将卷积结果与第二特征信息进行拼接，得到融合特征。
在本实施例中，上述执行主体可以将卷积结果与第二特征信息进行拼接，得到融合特征。将特征信息在通道维度上划分成第一特征信息和第二特征信息，仅对第一特征信息进行卷积，并与第二特征信息进行拼接，来得到融合特征，从而能够降低卷积计算量。
步骤207，将融合特征信息输入至全连接层，得到样本视频的预测类别。
步骤208，基于真实类别标签与预测类别的差异进行参数调整，得到视频识别模型。
在本实施例中，步骤207-208具体操作已在图1所示的实施例中步骤104-105进行了详细的介绍，在此不再赘述。
从图2中可以看出，与图1对应的实施例相比，本实施例中的视频识别模型训练方法突出了视频划分步骤、视频帧采样步骤和卷积融合步骤。由此，本实施例描述的方案，先根据视频长度对样本视频均匀划分，再对划分得到的样本视频片段均匀间隔采样，从而使得特征提取网络可以提取到样本视频的各个位置的特征信息。将特征信息在通道维度上划分成第一特征信息和第二特征信息，对第一特征信息进行卷积，并与第二特征信息进行拼接，来得到融合特征，从而能够降低卷积计算量。
进一步参考图3，其示出了可以实现本公开实施例的视频识别模 型训练方法的场景图。如图3所示，将样本视频均匀划分成4个样本视频片段(Snippets)，从每个样本视频片段中均匀间隔采样4帧视频帧。4个样本视频片段的4帧视频帧分别输入至CNN Layers，得到4个样本视频片段的特征信息。利用DSA Module对4个样本视频片段的特征信息进行卷积融合，得到的融合特征继续分别输入至CNN Layers进行处理。
进一步参考图4，其示出了视频识别模型的结构示意图。如图4所示，视频识别模型可以包括卷积层、多个残差层和全连接层，多个残差层可以间隔设置动态片段融合模块。具体地，视频识别模型包括Conv1、Res2、Res3、Res4、Res5和FC。样本视频的Snippets经过Conv1、Res2、Res3、Res4、Res5和FC的处理，得到样本视频的预测类别(属于各个预设类别的score)。在Res3和Res5内部分别设置两个动态融合模块。图4仅示出了Res3的结构，包括2个Res Block和2个DSA Block。Res5的结构与Res3的结构相同，图3未示出。
进一步参考图5，其示出了DSA Block的结构示意图。其中，图5示出了两种DSA Block。图5中的(a)示出了DSA Block(for TSM)，是2D DSA Block。图5中的(b)示出了DSA Block(for I3D)，是3D DSA Block。图5中的(c)示出了DSA Block(for TSM)和DSA Block(for I3D)中的DSA Module的结构示意图。DSA Module包括一个GAP和两个FC。特征信息在通道维度上划分成第一特征信息βC和第二特征信息(1-β)C。乘积βC×U×T×H×W输入至GAP，得到βC×U。βC×U输入至FC，得到βC×aU。βC×aU输入至FC，得到βC×L。βC×L与βC×U×T×H×W卷积，并与(1-β)C×U×T×H×W拼接。
进一步参考图6，其示出了根据本公开的视频识别方法的一个实施例的流程600。该视频识别方法包括以下步骤：
步骤601，获取待识别视频。
在本实施例中，视频识别方法的执行主体可以获取待识别视频。
步骤602，将待识别视频划分成多个待识别视频片段。
在本实施例中，上述执行主体可以将待识别视频划分成多个待识别视频片段。
这里，待识别视频的划分方式可以参见样本视频的划分方式，这里不再赘述。
在本实施例的一些可选的实现方式中，待识别视频的划分粒度大于用于训练视频识别模型的样本视频的划分粒度。用于训练视频识别模型的样本视频的数量巨大，减小样本视频的划分粒度可以缩短训练时间。而增加待识别视频的划分粒度可以提高识别精度。例如，对于10秒的样本视频，每隔2秒进行均匀划分，得到5个2秒的样本视频片段。而对于10秒的待识别视频，每隔1秒进行均匀划分，得到10个1秒的待识别视频片段。
步骤603，从待识别视频片段中采样部分待识别视频帧，并输入至视频识别模型，得到待识别视频的类别。
在本实施例中，从待识别视频片段中采样部分待识别视频帧，将其输入至视频识别模型进行预测，并对预测结果进行聚合，即可得到待识别视频的类别。
这里，待识别视频片段的采样方式可以参见样本视频片段的采样方式，这里不再赘述。视频识别模型可以用于视频分类，是利用图1-图2中任一实施方式提供的训练方法训练获得的，这里不再赘述。
本公开实施例提供的视频识别方法，提供了一种基于动态片段融合的高效视频识别方法，通过设计动态片段融合模块，使视频识别模型的卷积核能够在训练和推理中随着视频输入的不同而变化，从而提高了识别精度。视频识别模型采用动态卷积融合的识别方式，融合片段的卷积核参数能够随着输入的视频而变化，实现比单纯使用一种卷积核更为精准的时域感知，在提升识别精度的同时不会增加计算复杂度。尤其能够提升拥有更长更丰富的信息的长视频的识别精度。可用于中长视频分类、影视剧内容分类等。
进一步参考图7，作为对上述各图所示方法的实现，本公开提供了一种视频识别模型训练装置的一个实施例，该装置实施例与图1所示的方法实施例相对应，该装置具体可以应用于各种电子设备中。
如图7所示，本实施例的视频识别模型训练装置700可以包括：划分模块701、提取模块702、融合模块703、预测模块704和调整模块 705。其中，划分模块701，被配置成将样本视频划分成多个样本视频片段，其中，样本视频标注有真实类别标签；提取模块702，被配置成从样本视频片段中采样部分样本视频帧，并输入至特征提取网络，得到样本视频片段的特征信息；融合模块703，被配置成利用动态片段融合模块对特征信息进行卷积融合，得到融合特征信息，其中，动态片段融合模块的卷积核随视频输入的不同而变化；预测模块704，被配置成将融合特征信息输入至全连接层，得到样本视频的预测类别；调整模块705，被配置成基于真实类别标签与预测类别的差异进行参数调整，得到视频识别模型。
在本实施例中，视频识别模型训练装置700中：划分模块701、提取模块702、融合模块703、预测模块704和调整模块705的具体处理及其所带来的技术效果可分别参考图1对应实施例中的步骤101-105的相关说明，在此不再赘述。
在本实施例的一些可选的实现方式中，融合模块703包括：划分子模块，被配置成将特征信息在通道维度上划分成第一特征信息和第二特征信息；确定子模块，被配置成利用卷积核生成分支网络确定样本视频对应的卷积核；卷积子模块，被配置成利用样本视频对应的卷积核对第一特征信息进行卷积，得到卷积结果；拼接子模块，被配置成将卷积结果与第二特征信息进行拼接，得到融合特征。
在本实施例的一些可选的实现方式中，划分子模块进一步被配置成：按照预设超参数β将特征信息在通道维度上划分成第一特征信息和第二特征信息，其中，第一特征信息的通道维度是βC，第二特征信息的通道维度是(1-β)C，C是特征信息的通道维度。
在本实施例的一些可选的实现方式中，确定子模块进一步被配置成：计算第一特征信息的通道维度βC、样本视频的片段数、样本视频片段的采样帧数、样本视频帧的高和宽的乘积；将乘积输入至卷积核生成分支网络，得到样本视频对应的卷积核。
在本实施例的一些可选的实现方式中，卷积核生成分支网络包括一个全局平均池化层和两个全连接层。
在本实施例的一些可选的实现方式中，视频识别模型包括多个残差层，多个残差层间隔设置至少一个动态片段融合模块。
在本实施例的一些可选的实现方式中，划分模块701进一步被配置成：根据视频长度对样本视频进行均匀划分，得到多个样本视频片段；以及提取模块1002进一步被配置成：对样本视频片段进行均匀间隔采样，得到部分样本视频帧。
在本实施例的一些可选的实现方式中，调整模块705进一步被配置成：基于真实类别标签和预测类别计算交叉熵损失；使用随机梯度下降对交叉熵损失进行优化，不断进行参数更新，直至交叉熵损失收敛，得到视频识别模型。
进一步参考图8，作为对上述各图所示方法的实现，本公开提供了一种视频识别装置的一个实施例，该装置实施例与图6所示的方法实施例相对应，该装置具体可以应用于各种电子设备中。
如图8所示，本实施例的视频识别装置800可以包括：获取模块801、划分模块802和识别模块803。其中，获取模块801，被配置成获取待识别视频；划分模块802，被配置成将待识别视频划分成多个待识别视频片段；识别模块803，被配置成从待识别视频片段中采样部分待识别视频帧，并输入至视频识别模型，得到待识别视频的类别，其中，视频识别模型根据图1-图2中任一实施例描述的训练方法训练获得。
在本实施例中，视频识别装置800中：获取模块801、划分模块802和识别模块803的具体处理及其所带来的技术效果可分别参考图6对应实施例中的步骤601-603的相关说明，在此不再赘述。
在本实施例的一些可选的实现方式中，待识别视频的划分粒度大于用于训练视频识别模型的样本视频的划分粒度。
本公开的技术方案中，所涉及的用户个人信息的获取，存储和应用等，均符合相关法律法规的规定，且不违背公序良俗。
根据本公开的实施例，本公开还提供了一种电子设备、一种可读存储介质和一种计算机程序产品。
图9示出了可以用来实施本公开的实施例的示例电子设备900的示意性框图。电子设备旨在表示各种形式的数字计算机，诸如，膝上型计算机、台式计算机、工作台、个人数字助理、服务器、刀片式服务器、大型计算机、和其它适合的计算机。电子设备还可以表示各种 形式的移动装置，诸如，个人数字处理、蜂窝电话、智能电话、可穿戴设备和其它类似的计算装置。本文所示的部件、它们的连接和关系、以及它们的功能仅仅作为示例，并且不意在限制本文中描述的和/或者要求的本公开的实现。
如图9所示，设备900包括计算单元901，其可以根据存储在只读存储器(ROM)902中的计算机程序或者从存储单元908加载到随机访问存储器(RAM)903中的计算机程序，来执行各种适当的动作和处理。在RAM 903中，还可存储设备900操作所需的各种程序和数据。计算单元901、ROM 902以及RAM 903通过总线904彼此相连。输入/输出(I/O)接口905也连接至总线904。
设备900中的多个部件连接至I/O接口905，包括：输入单元906，例如键盘、鼠标等；输出单元907，例如各种类型的显示器、扬声器等；存储单元908，例如磁盘、光盘等；以及通信单元909，例如网卡、调制解调器、无线通信收发机等。通信单元909允许设备900通过诸如因特网的计算机网络和/或各种电信网络与其他设备交换信息/数据。
计算单元901可以是各种具有处理和计算能力的通用和/或专用处理组件。计算单元901的一些示例包括但不限于中央处理单元(CPU)、图形处理单元(GPU)、各种专用的人工智能(AI)计算芯片、各种运行机器学习模型算法的计算单元、数字信号处理器(DSP)、以及任何适当的处理器、控制器、微控制器等。计算单元901执行上文所描述的各个方法和处理，例如视频识别模型训练方法。例如，在一些实施例中，视频识别模型训练方法可被实现为计算机软件程序，其被有形地包含于机器可读介质，例如存储单元908。在一些实施例中，计算机程序的部分或者全部可以经由ROM 902和/或通信单元909而被载入和/或安装到设备900上。当计算机程序加载到RAM 903并由计算单元901执行时，可以执行上文描述的视频识别模型训练方法的一个或多个步骤。备选地，在其他实施例中，计算单元901可以通过其他任何适当的方式(例如，借助于固件)而被配置为执行视频识别模型训练方法。
本文中以上描述的系统和技术的各种实施方式可以在数字电子电 路系统、集成电路系统、场可编程门阵列(FPGA)、专用集成电路(ASIC)、专用标准产品(ASSP)、芯片上系统的系统(SOC)、负载可编程逻辑设备(CPLD)、计算机硬件、固件、软件、和/或它们的组合中实现。这些各种实施方式可以包括：实施在一个或者多个计算机程序中，该一个或者多个计算机程序可在包括至少一个可编程处理器的可编程系统上执行和/或解释，该可编程处理器可以是专用或者通用可编程处理器，可以从存储系统、至少一个输入装置、和至少一个输出装置接收数据和指令，并且将数据和指令传输至该存储系统、该至少一个输入装置、和该至少一个输出装置。
用于实施本公开的方法的程序代码可以采用一个或多个编程语言的任何组合来编写。这些程序代码可以提供给通用计算机、专用计算机或其他可编程数据处理装置的处理器或控制器，使得程序代码当由处理器或控制器执行时使流程图和/或框图中所规定的功能/操作被实施。程序代码可以完全在机器上执行、部分地在机器上执行，作为独立软件包部分地在机器上执行且部分地在远程机器上执行或完全在远程机器或服务器上执行。
在本公开的上下文中，机器可读介质可以是有形的介质，其可以包含或存储以供指令执行系统、装置或设备使用或与指令执行系统、装置或设备结合地使用的程序。机器可读介质可以是机器可读信号介质或机器可读储存介质。机器可读介质可以包括但不限于电子的、磁性的、光学的、电磁的、红外的、或半导体系统、装置或设备，或者上述内容的任何合适组合。机器可读存储介质的更具体示例会包括基于一个或多个线的电气连接、便携式计算机盘、硬盘、随机存取存储器(RAM)、只读存储器(ROM)、可擦除可编程只读存储器(EPROM或快闪存储器)、光纤、便捷式紧凑盘只读存储器(CD-ROM)、光学储存设备、磁储存设备、或上述内容的任何合适组合。
为了提供与用户的交互，可以在计算机上实施此处描述的系统和技术，该计算机具有：用于向用户显示信息的显示装置(例如，CRT(阴极射线管)或者LCD(液晶显示器)监视器)；以及键盘和指向装置(例如，鼠标或者轨迹球)，用户可以通过该键盘和该指向装置 来将输入提供给计算机。其它种类的装置还可以用于提供与用户的交互；例如，提供给用户的反馈可以是任何形式的传感反馈(例如，视觉反馈、听觉反馈、或者触觉反馈)；并且可以用任何形式(包括声输入、语音输入或者、触觉输入)来接收来自用户的输入。
可以将此处描述的系统和技术实施在包括后台部件的计算系统(例如，作为数据服务器)、或者包括中间件部件的计算系统(例如，应用服务器)、或者包括前端部件的计算系统(例如，具有图形用户界面或者网络浏览器的用户计算机，用户可以通过该图形用户界面或者该网络浏览器来与此处描述的系统和技术的实施方式交互)、或者包括这种后台部件、中间件部件、或者前端部件的任何组合的计算系统中。可以通过任何形式或者介质的数字数据通信(例如，通信网络)来将系统的部件相互连接。通信网络的示例包括：局域网(LAN)、广域网(WAN)和互联网。
计算机系统可以包括客户端和服务器。客户端和服务器一般远离彼此并且通常通过通信网络进行交互。通过在相应的计算机上运行并且彼此具有客户端-服务器关系的计算机程序来产生客户端和服务器的关系。服务器可以是云服务器，也可以是分布式系统的服务器，或者是结合了区块链的服务器。
应该理解，可以使用上面所示的各种形式的流程，重新排序、增加或删除步骤。例如，本公开中记载的各步骤可以并行地执行也可以顺序地执行也可以不同的次序执行，只要能够实现本公开提供的技术方案所期望的结果，本文在此不进行限制。
上述具体实施方式，并不构成对本公开保护范围的限制。本领域技术人员应该明白的是，根据设计要求和其他因素，可以进行各种修改、组合、子组合和替代。任何在本公开的精神和原则之内所作的修改、等同替换和改进等，均应包含在本公开保护范围之内。

Claims (23)

  1. 一种视频识别模型训练方法，包括：
    将样本视频划分成多个样本视频片段，其中，所述样本视频标注有真实类别标签；
    从所述样本视频片段中采样部分样本视频帧，并输入至特征提取网络，得到所述样本视频片段的特征信息；
    利用动态片段融合模块对所述特征信息进行卷积融合，得到融合特征信息，其中，所述动态片段融合模块的卷积核随视频输入的不同而变化；
    将所述融合特征信息输入至全连接层，得到所述样本视频的预测类别；
    基于所述真实类别标签与所述预测类别的差异进行参数调整，得到所述视频识别模型。
  2. 根据权利要求1所述的方法，其中，所述利用动态片段融合模块对所述特征信息进行卷积融合，得到融合特征信息，包括：
    将所述特征信息在通道维度上划分成第一特征信息和第二特征信息；
    利用卷积核生成分支网络确定所述样本视频对应的卷积核；
    利用所述样本视频对应的卷积核对所述第一特征信息进行卷积，得到卷积结果；
    将所述卷积结果与所述第二特征信息进行拼接，得到所述融合特征。
  3. 根据权利要求2所述的方法，其中，所述将所述特征信息在通道维度上划分成第一特征信息和第二特征信息，包括：
    按照预设超参数β将所述特征信息在通道维度上划分成所述第一特征信息和所述第二特征信息，其中，所述第一特征信息的通道维度是βC，所述第二特征信息的通道维度是(1-β)C，C是所述特征信息的通道维度。
  4. 根据权利要求3所述的方法，其中，所述利用卷积核生成分支网络确定所述样本视频对应的卷积核，包括：
    计算所述第一特征信息的通道维度βC、所述样本视频的片段数、所述样本视频片段的采样帧数、所述样本视频帧的高和宽的乘积；
    将所述乘积输入至所述卷积核生成分支网络，得到所述样本视频对应的卷积核。
  5. 根据权利要求2-4中任一项所述的方法，其中，所述卷积核生成分支网络包括一个全局平均池化层和两个全连接层。
  6. 根据权利要求1-5中任一项所述的方法，其中，所述动态片段融合模块包含至少一个动态片段融合模块，所述视频识别模型包括所述至少一个动态片段融合模块及多个残差层，所述多个残差层间隔设置所述至少一个动态片段融合模块。
  7. 根据权利要求1-6中任一项所述的方法，其中，所述将样本视频划分成多个样本视频片段，包括：
    根据视频长度对所述样本视频进行均匀划分，得到所述多个样本视频片段；以及
    所述从所述样本视频片段中采样部分样本视频帧，包括：
    对所述样本视频片段进行均匀间隔采样，得到所述部分样本视频帧。
  8. 根据权利要求1-7中任一项所述的方法，其中，所述基于所述真实类别标签与所述预测类别的差异进行参数调整，得到所述视频识别模型，包括：
    基于所述真实类别标签和所述预测类别计算交叉熵损失；
    使用随机梯度下降对所述交叉熵损失进行优化，不断进行参数更新，直至所述交叉熵损失收敛，得到所述视频识别模型。
  9. 一种视频识别方法，包括：
    获取待识别视频；
    将所述待识别视频划分成多个待识别视频片段；
    从所述待识别视频片段中采样部分待识别视频帧，并输入至视频识别模型，得到所述待识别视频的类别，其中，所述视频识别模型根据权利要求1-8中任一项的训练方法训练获得。
  10. 根据权利要求9所述的方法，其中，所述待识别视频的划分粒度大于用于训练所述视频识别模型的样本视频的划分粒度。
  11. 一种视频识别模型训练装置，包括：
    划分模块，被配置成将样本视频划分成多个样本视频片段，其中，所述样本视频标注有真实类别标签；
    提取模块，被配置成从所述样本视频片段中采样部分样本视频帧，并输入至特征提取网络，得到所述样本视频片段的特征信息；
    融合模块，被配置成利用动态片段融合模块对所述特征信息进行卷积融合，得到融合特征信息，其中，所述动态片段融合模块的卷积核随视频输入的不同而变化；
    预测模块，被配置成将所述融合特征信息输入至全连接层，得到所述样本视频的预测类别；
    调整模块，被配置成基于所述真实类别标签与所述预测类别的差异进行参数调整，得到所述视频识别模型。
  12. 根据权利要求11所述的装置，其中，所述融合模块包括：
    划分子模块，被配置成将所述特征信息在通道维度上划分成第一特征信息和第二特征信息；
    确定子模块，被配置成利用卷积核生成分支网络确定所述样本视频对应的卷积核；
    卷积子模块，被配置成利用所述样本视频对应的卷积核对所述第一特征信息进行卷积，得到卷积结果；
    拼接子模块，被配置成将所述卷积结果与所述第二特征信息进行拼接，得到所述融合特征。
  13. 根据权利要求12所述的装置，其中，所述划分子模块进一步被配置成：
    按照预设超参数β将所述特征信息在通道维度上划分成所述第一特征信息和所述第二特征信息，其中，所述第一特征信息的通道维度是βC，所述第二特征信息的通道维度是(1-β)C，C是所述特征信息的通道维度。
  14. 根据权利要求13所述的装置，其中，所述确定子模块进一步被配置成：
    计算所述第一特征信息的通道维度βC、所述样本视频的片段数、所述样本视频片段的采样帧数、所述样本视频帧的高和宽的乘积；
    将所述乘积输入至所述卷积核生成分支网络，得到所述样本视频对应的卷积核。
  15. 根据权利要求12-14中任一项所述的装置，其中，所述卷积核生成分支网络包括一个全局平均池化层和两个全连接层。
  16. 根据权利要求11-15中任一项所述的装置，其中，所述动态片段融合模块包含至少一个动态片段融合模块，所述视频识别模型包括所述至少一个动态片段融合模块及多个残差层，所述多个残差层间隔设置所述至少一个动态片段融合模块。
  17. 根据权利要求11-16中任一项所述的装置，其中，所述划分模块进一步被配置成：
    根据视频长度对所述样本视频进行均匀划分，得到所述多个样本视频片段；以及
    所述提取模块进一步被配置成：
    对所述样本视频片段进行均匀间隔采样，得到所述部分样本视频帧。
  18. 根据权利要求11-17中任一项所述的装置，其中，所述调整模块进一步被配置成：
    基于所述真实类别标签和所述预测类别计算交叉熵损失；
    使用随机梯度下降对所述交叉熵损失进行优化，不断进行参数更新，直至所述交叉熵损失收敛，得到所述视频识别模型。
  19. 一种视频识别装置，包括：
    获取模块，被配置成获取待识别视频；
    划分模块，被配置成将所述待识别视频划分成多个待识别视频片段；
    识别模块，被配置成从所述待识别视频片段中采样部分待识别视频帧，并输入至视频识别模型，得到所述待识别视频的类别，其中，所述视频识别模型根据权利要求1-8中任一项的训练方法训练获得。
  20. 根据权利要求19所述的装置，其中，所述待识别视频的划分粒度大于用于训练所述视频识别模型的样本视频的划分粒度。
  21. 一种电子设备，包括：
    至少一个处理器；以及
    与所述至少一个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至少一个处理器执行的指令，所述指令被所述至少一个处理器执行，以使所述至少一个处理器能够执行权利要求1-8中任一项所述的方法，或者执行权利要求9或10所述的方法。
  22. 一种存储有计算机指令的非瞬时计算机可读存储介质，所述计算机指令用于使所述计算机执行权利要求1-8中任一项所述的方法，或者执行权利要求9或10所述的方法。
  23. 一种计算机程序产品，包括计算机程序，所述计算机程序在被处理器执行时实现根据权利要求1-8中任一项所述的方法，或者执行权利要求9或10所述的方法。
PCT/CN2022/075153 2021-05-28 2022-01-30 视频识别模型训练方法、装置、设备以及存储介质 WO2022247344A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
JP2022563231A JP7417759B2 (ja) 2021-05-28 2022-01-30 ビデオ認識モデルをトレーニングする方法、装置、電子機器、記憶媒体およびコンピュータプログラム
US17/983,208 US20230069197A1 (en) 2021-05-28 2022-11-08 Method, apparatus, device and storage medium for training video recognition model

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202110589375.6 2021-05-28
CN202110589375.6A CN113326767A (zh) 2021-05-28 2021-05-28 视频识别模型训练方法、装置、设备以及存储介质

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US17/983,208 Continuation US20230069197A1 (en) 2021-05-28 2022-11-08 Method, apparatus, device and storage medium for training video recognition model

Publications (1)

Publication Number Publication Date
WO2022247344A1 true WO2022247344A1 (zh) 2022-12-01

Family

ID=77422144

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/075153 WO2022247344A1 (zh) 2021-05-28 2022-01-30 视频识别模型训练方法、装置、设备以及存储介质

Country Status (4)

Country Link
US (1) US20230069197A1 (zh)
JP (1) JP7417759B2 (zh)
CN (1) CN113326767A (zh)
WO (1) WO2022247344A1 (zh)

Families Citing this family (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113326767A (zh) * 2021-05-28 2021-08-31 北京百度网讯科技有限公司 视频识别模型训练方法、装置、设备以及存储介质
CN113741459B (zh) * 2021-09-03 2024-06-21 阿波罗智能技术(北京)有限公司 确定训练样本的方法和自动驾驶模型的训练方法、装置
CN113487247B (zh) * 2021-09-06 2022-02-01 阿里巴巴（中国）有限公司 数字化生产管理系统、视频处理方法、设备及存储介质
CN113963287A (zh) * 2021-09-15 2022-01-21 北京百度网讯科技有限公司 评分模型获取及视频识别方法、装置及存储介质
CN114218438B (zh) * 2021-12-23 2023-03-21 北京百度网讯科技有限公司 视频数据处理方法、装置、电子设备和计算机存储介质
CN114359811A (zh) * 2022-01-11 2022-04-15 北京百度网讯科技有限公司 数据鉴伪方法、装置、电子设备以及存储介质
CN114419508A (zh) * 2022-01-19 2022-04-29 北京百度网讯科技有限公司 识别方法、训练方法、装置、设备及存储介质
CN114882334B (zh) * 2022-04-29 2023-04-28 北京百度网讯科技有限公司 用于生成预训练模型的方法、模型训练方法及装置
CN117011740A (zh) * 2022-10-20 2023-11-07 腾讯科技（深圳）有限公司 视频检测方法和装置、存储介质及电子设备
CN116132752B (zh) * 2023-04-13 2023-12-08 北京百度网讯科技有限公司 视频对照组构造、模型训练、视频打分方法、装置及设备
CN116493392B (zh) * 2023-06-09 2023-12-15 北京中超伟业信息安全技术股份有限公司 一种纸介质碳化方法及系统
CN117612072B (zh) * 2024-01-23 2024-04-19 中国科学技术大学 一种基于动态时空图的视频理解方法

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190297326A1 (en) * 2018-03-21 2019-09-26 Nvidia Corporation Video prediction using spatially displaced convolution
CN111008280A (zh) * 2019-12-04 2020-04-14 北京百度网讯科技有限公司 一种视频分类方法、装置、设备和存储介质
CN111241985A (zh) * 2020-01-08 2020-06-05 腾讯科技（深圳）有限公司 一种视频内容识别方法、装置、存储介质、以及电子设备
CN113326767A (zh) * 2021-05-28 2021-08-31 北京百度网讯科技有限公司 视频识别模型训练方法、装置、设备以及存储介质

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112232407B (zh) * 2020-10-15 2023-08-18 杭州迪英加科技有限公司 病理图像样本的神经网络模型训练方法、装置

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190297326A1 (en) * 2018-03-21 2019-09-26 Nvidia Corporation Video prediction using spatially displaced convolution
CN111008280A (zh) * 2019-12-04 2020-04-14 北京百度网讯科技有限公司 一种视频分类方法、装置、设备和存储介质
CN111241985A (zh) * 2020-01-08 2020-06-05 腾讯科技（深圳）有限公司 一种视频内容识别方法、装置、存储介质、以及电子设备
CN113326767A (zh) * 2021-05-28 2021-08-31 北京百度网讯科技有限公司 视频识别模型训练方法、装置、设备以及存储介质

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
WENHAO WU; YUXIANG ZHAO; YANWU XU; XIAO TAN; DONGLIANG HE; ZHIKANG ZOU; JIN YE; YINGYING LI; MINGDE YAO; ZICHAO DONG; YIFENG SHI: "DSANet: Dynamic Segment Aggregation Network for Video-Level Representation Learning", ARXIV.ORG, CORNELL UNIVERSITY LIBRARY, 201 OLIN LIBRARY CORNELL UNIVERSITY ITHACA, NY 14853, 7 July 2021 (2021-07-07), 201 Olin Library Cornell University Ithaca, NY 14853, XP091005237 *

Also Published As

Publication number Publication date
CN113326767A (zh) 2021-08-31
JP7417759B2 (ja) 2024-01-18
JP2023531132A (ja) 2023-07-21
US20230069197A1 (en) 2023-03-02

Similar Documents

Publication Publication Date Title
WO2022247344A1 (zh) 视频识别模型训练方法、装置、设备以及存储介质
CN111061881A (zh) 文本分类方法、设备及存储介质
KR20210124111A (ko) 모델을 훈련하기 위한 방법, 장치, 기기, 매체 및 프로그램 제품
JP2022135991A (ja) クロスモーダル検索モデルのトレーニング方法、装置、機器、および記憶媒体
US20230090590A1 (en) Speech recognition and codec method and apparatus, electronic device and storage medium
US11816891B2 (en) Video recognition method and apparatus, electronic device and storage medium
US20230047980A1 (en) Method of training deep learning model and method of processing natural language
US20230022677A1 (en) Document processing
WO2023015939A1 (zh) 用于文本检测的深度学习模型训练方法及文本检测方法
CN113326852A (zh) 模型训练方法、装置、设备、存储介质及程序产品
CN113657483A (zh) 模型训练方法、目标检测方法、装置、设备以及存储介质
CN114020950B (zh) 图像检索模型的训练方法、装置、设备以及存储介质
US20230306081A1 (en) Method for training a point cloud processing model, method for performing instance segmentation on point cloud, and electronic device
US20220374678A1 (en) Method for determining pre-training model, electronic device and storage medium
WO2023005253A1 (zh) 文本识别模型框架的训练方法、装置及系统
US20230066021A1 (en) Object detection
US20220398834A1 (en) Method and apparatus for transfer learning
CN112949818A (zh) 模型蒸馏方法、装置、设备以及存储介质
CN112949433B (zh) 视频分类模型的生成方法、装置、设备和存储介质
CN113360672B (zh) 用于生成知识图谱的方法、装置、设备、介质和产品
CN114926322A (zh) 图像生成方法、装置、电子设备和存储介质
CN114707638A (zh) 模型训练、对象识别方法及装置、设备、介质和产品
CN113361621A (zh) 用于训练模型的方法和装置
CN113313049A (zh) 超参数的确定方法、装置、设备、存储介质以及计算机程序产品
CN113657466A (zh) 预训练模型的生成方法、装置、电子设备和存储介质

Legal Events

Date Code Title Description
ENP Entry into the national phase

Ref document number: 2022563231

Country of ref document: JP

Kind code of ref document: A

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

Ref document number: 22810080

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 22810080

Country of ref document: EP

Kind code of ref document: A1