Video analysis method, device and medium combining active learning and noise screening

Abstract

The application discloses a video analysis method and device combining active learning and noise screening and a medium, and the method comprises the following steps: acquiring initial images and adding noises with different intensities; performing target detection on each image in an image sequence by using a target detection network; calculating the maximum confidence and the uncertainty score of each initial image, and dividing the initial images into target samples and manually labeled samples according to the calculation results; performing cyclic iteration training on the labeled samples by using the target detection network, and recording the sample loss in the cyclic iteration training process; selecting target samples with sample loss less than a preset loss threshold as pseudo-label samples, training the target detection network in combination with the pseudo-label samples and the manually labeled samples, and performing detection analysis on specified video images by using the trained target detection network. The application is based on the active learning integrated pseudo-label learning and the deep learning training noise sample screening technology, and can improve the analysis efficiency and accuracy of video images.

Description

Technical Field

[0001] This invention relates to the field of active learning technology, and in particular to a video analysis method, apparatus, and medium that combines active learning and noise removal. Background Technology

[0002] In the field of video analytics, continuous real-time analysis of surveillance footage is required around the clock, resulting in the collection of a large number of unlabeled images. Although labeling all images individually can improve model accuracy, introducing a large amount of manual annotation is too time-consuming and labor-intensive, inevitably causing cost issues.

[0003] The core idea of ​​active learning is to use models and machine learning methods to select higher-value samples from unlabeled samples for manual annotation. This allows for the rapid and cost-effective replenishment of the model's training set, enabling continuous model iteration and optimization. The strategy for selecting training samples is called a query strategy. Existing methods typically calculate indicators such as sample confidence and information entropy, using thresholds to categorize difficult and easy examples, and assigning difficult examples to human annotators. In other words, selecting higher-value samples for manual annotation reduces the workload while ensuring the representativeness of the labeled samples. However, when faced with massive amounts of samples, the manual workload remains enormous, making efficient annotation impossible. Summary of the Invention

[0004] This invention provides a video analysis method, apparatus, computer device, and storage medium that combines active learning and noise removal, aiming to improve the efficiency and accuracy of video image analysis.

[0005] In a first aspect, embodiments of the present invention provide a video analysis method combining active learning and noise removal, comprising:

[0006] Multiple unlabeled images are obtained as initial images, and noise of different intensities is added to each initial image to obtain an image sequence containing the initial images and multiple noisy images;

[0007] The target detection network is used to perform target detection on each image in the image sequence to obtain the target detection result for each image;

[0008] The maximum confidence score and uncertainty score of each initial image are calculated based on the target detection results, and the initial images are divided into target samples and manually labeled samples according to the calculation results;

[0009] The labeled samples are trained iteratively using the target detection network, and the sample loss during the iterative training process is recorded.

[0010] Target samples with a loss less than a preset loss threshold are selected as pseudo-label samples. The target detection network is trained by combining the pseudo-label samples and manually labeled samples. The trained target detection network is then used to detect and analyze specified video images.

[0011] Secondly, embodiments of the present invention provide a video analysis apparatus combining active learning and noise removal, comprising:

[0012] An image acquisition unit is used to acquire multiple unlabeled images as initial images, and add noise of different intensities to each initial image to obtain an image sequence containing the initial image and multiple noisy images;

[0013] The target detection unit is used to perform target detection on each image in the image sequence using a target detection network, and obtain the target detection result for each image;

[0014] The sample screening unit is used to calculate the maximum confidence score and uncertainty score of each initial image based on the target detection results, and to divide the initial images into target samples and manually labeled samples according to the calculation results;

[0015] The iterative training unit is used to perform iterative training on the labeled samples using the target detection network and record the sample loss during the iterative training process.

[0016] The video analysis unit is used to select target samples with sample loss less than a preset loss threshold as pseudo-label samples, combine the pseudo-label samples and manually labeled samples to train the target detection network, and use the trained target detection network to detect and analyze specified video images.

[0017] Thirdly, embodiments of the present invention provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the video analysis method combining active learning and noise removal as described in the first aspect.

[0018] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the video analysis method combining active learning and noise removal as described in the first aspect.

[0019] This invention provides a video analysis method, apparatus, computer device, and storage medium that combines active learning and noise removal. The method includes: acquiring multiple unlabeled images as initial images, and adding noise of different intensities to each initial image to obtain an image sequence containing the initial images and multiple noisy images; performing target detection on each image in the image sequence using a target detection network to obtain a target detection result for each image; calculating the maximum confidence score and uncertainty score of each initial image based on the target detection results, and dividing the initial images into target samples and manually labeled samples according to the calculation results; performing iterative training on the labeled samples using the target detection network and recording the sample loss during the iterative training process; selecting target samples with sample loss less than a preset loss threshold as pseudo-label samples, training the target detection network using the pseudo-label samples and manually labeled samples, and using the trained target detection network to detect and analyze specified video images. This invention integrates pseudo-label learning and deep learning training noise sample screening techniques, and performs secondary screening to address the noise issues present in pseudo-label learning. The selected samples are then used to train the target detection network, which in turn is used to detect and analyze video images. This improves the efficiency and accuracy of video image analysis. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 A flowchart illustrating a video analysis method combining active learning and noise removal, provided as an embodiment of the present invention;

[0022] Figure 2 A schematic diagram of a sub-process of a video analysis method combining active learning and noise removal provided in an embodiment of the present invention;

[0023] Figure 3 A network flowchart of a video analysis method combining active learning and noise removal is provided for an embodiment of the present invention;

[0024] Figure 4 A schematic block diagram of a video analysis device combining active learning and noise removal, provided for an embodiment of the present invention;

[0025] Figure 5 This is a schematic block diagram of a video analysis device that combines active learning and noise removal, provided as an embodiment of the present invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0028] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0029] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0030] Please see below. Figure 1 , Figure 1 The flowchart of a video analysis method combining active learning and noise removal provided in an embodiment of the present invention specifically includes steps S101 to S105.

[0031] S101. Obtain multiple unlabeled images as initial images, and add noise of different intensities to each initial image to obtain an image sequence containing the initial image and multiple noisy images;

[0032] S102. Target detection is performed on each image in the image sequence using a target detection network to obtain the target detection result for each image;

[0033] S103. Calculate the maximum confidence score and uncertainty score of each initial image based on the target detection results, and divide the initial image into target samples and manually labeled samples according to the calculation results;

[0034] S104. The labeled samples are trained iteratively using the target detection network, and the sample loss during the iterative training process is recorded.

[0035] S105. Select target samples with sample loss less than a preset loss threshold as pseudo-label samples, combine the pseudo-label samples and manually labeled samples to train the target detection network, and use the trained target detection network to detect and analyze the specified video images.

[0036] In this embodiment, noise is first added to the unlabeled initial image to form an image sequence. This image sequence is then input into an object detection network, which outputs the corresponding object detection results. Based on these results, the maximum confidence score and uncertainty score of the initial image can be calculated. Furthermore, based on these scores, the initial image is divided into target samples and manually labeled samples (i.e., samples requiring manual labeling). The target samples are then iteratively trained to select suitable pseudo-labeled samples. Subsequently, the pseudo-labeled samples and manually labeled samples are combined into a training set to train and optimize the object detection network. The optimized network then performs detection and analysis on specified video images.

[0037] This embodiment, based on the concept of active learning, integrates pseudo-label learning and deep learning training noise sample screening techniques. It can more accurately identify difficult samples requiring manual annotation, while also performing a secondary screening to address noise issues inherent in pseudo-label learning. In other words, this embodiment automatically and accurately filters samples that the model can accurately identify, using the results as labels or coarse labels. Simultaneously, based on the calculation of uncertainty scores, it mines difficult examples from massive datasets for annotation personnel. The results of both methods are then added to the training samples. This significantly reduces the workload of manual annotation while rapidly accumulating the labeled samples needed by the model, facilitating continuous model operation and rapid iterative training, thereby improving the efficiency and accuracy of video image analysis.

[0038] Furthermore, the target detection network described in this embodiment can be a widely used target detection network, and the target detection network can be constructed using labeled images to obtain a target detection network suitable for this embodiment.

[0039] In one embodiment, step S101 includes:

[0040] The image sequence [X0, X1, X2, ... X] is obtained by adding noise according to the following formula. i ...X n ]:

[0041] X i =X0+ε i

[0042] In the formula, noise ε i ∈N(0,σi 2 ) follows a Gaussian distribution, σ i X represents the standard deviation, X0 represents the initial image, X i Let i represent the i-th noisy image.

[0043] In this embodiment, by adding noise of different intensities to the initial image according to the above noise addition formula, multiple noisy images corresponding to the initial image can be obtained, and thus the image sequence can be constructed.

[0044] In one embodiment, step S102 includes:

[0045] The YOLOv5 object detection network is used to perform object detection on each image in the image sequence, and the YOLOv5 object detection network outputs the location information b and confidence p of the object detection box of the business object in each image.

[0046] The YOLOv5 network can predict and output the location coordinates of the corresponding bounding box for each image in an image sequence (b i ), target confidence level (p i ) and category (c i Of course, other target detection networks can be used in other application scenarios, and the choice of target detection network does not affect the analysis results of this embodiment.

[0047] In one embodiment, step S103 includes:

[0048] Initial images with a maximum confidence score greater than a preset confidence score and / or an uncertainty score less than a preset score threshold are selected as target samples, and the remaining initial images are selected as manually labeled samples.

[0049] When dividing the initial image into target samples and manually labeled samples based on the calculation results, specifically, initial images with a maximum confidence score greater than a preset confidence score and / or an uncertainty score less than a preset score threshold are divided into target samples, and initial images with a maximum confidence score less than or equal to a preset confidence score and / or an uncertainty score greater than or equal to a preset score threshold are divided into manually labeled samples. The target samples will be filtered to obtain pseudo-label samples after subsequent processing steps, and the remaining samples in the target samples that are not filtered will be manually labeled together with the manually labeled samples.

[0050] In one embodiment, step S103 further includes:

[0051] For each initial image, the Intersection over Union (IoU) of the target detection box in the initial image and the target detection boxes in all corresponding noisy images is calculated, and the target detection box in the noisy image with the maximum IoU is selected as the matching box.

[0052] Furthermore, step S103 also includes:

[0053] For each initial image, the maximum confidence α of the initial image is calculated according to the following formula:

[0054]

[0055] In the formula, p0 represents the confidence level of the target detection box in the initial image. This indicates the confidence level of the target detection box;

[0056] And the positional uncertainty β of the initial image is calculated according to the following formula:

[0057]

[0058] In the formula, b0 represents the position information of the target detection box in the initial image. This represents the position information of the matching box, where IoU represents the relationship between b0 and... The intersection and union ratio;

[0059] Traverse all noisy images, calculate the average of the maximum confidence α and the average of the location uncertainty β of the initial image, and perform a weighted sum of the average of the maximum confidence α and the average of the location uncertainty β to obtain the uncertainty score of the initial image.

[0060] In this embodiment, for the noisy image X i Search results b i and p i The detection results are compared with those of the initial image, and the Intersection over Union (IoU) between the target detection boxes in the initial image and the noisy image is calculated. The target detection boxes in the initial image are traversed, and the box with the largest IoU in the noisy image is selected as the matching box.

[0061] The confidence uncertainty of the initial image is then calculated based on the matching bounding box. and location uncertainty Here, if the target detection box in the noisy image cannot be matched with the target detection box in the initial image, that is, the original target cannot be detected after the addition of noise, then α = 1, β = 1.

[0062] Next, iterate through all noise levels and all target detection boxes, calculate the average of the confidence uncertainty α and the location uncertainty β, and then obtain the uncertainty score S by weighted summation, for example... Samples with the highest confidence level of positional uncertainty in the initial image are selected. The resulting uncertainty score is calculated. If the score is too high, the sample is classified as one that requires manual annotation. If the score is too low, the sample is classified as the target sample.

[0063] In one embodiment, such as Figure 2 As shown, step S104 includes steps S201 to S203.

[0064] S201. Set the number of loops and set the initial learning rate for iterative training.

[0065] S202. An exponential decay learning rate strategy is adopted to control the decrease of the learning rate in each training process.

[0066] S203. When the target detection network is in an overfitting state, the current training ends, and the initial learning rate is restored at the start of the next training session to complete the iterative training.

[0067] In this embodiment, if the target samples obtained from the segmentation are directly analyzed using an object detection network, the resulting analysis results will be used as pseudo-labels, generating noise and affecting training accuracy. Therefore, an iterative training method is used to filter out noisy labels. Specifically, a training learning rate of l0 is initially set, and the learning rate is gradually decreased during training until the training converges. Once the object detection network overfits, the learning rate is restored to l0, and this process is repeated multiple times.

[0068] Furthermore, the loss of all target samples during iterative training is recorded. Since the loss of noisy target samples decreases more drastically with decreasing learning rate, while the loss of clean (i.e., noise-free or low-intensity) target samples remains low and relatively stable from the start of training, the K largest loss samples are selected as noisy samples, and the rest as clean samples. Noisy samples are categorized as manually labeled samples, and clean samples are categorized as pseudo-labeled samples. Then, pseudo-labeled samples are directly added to the labeled samples (i.e., the training set), using pseudo-labels as training labels. Simultaneously, manually labeled samples are assigned to annotators, and the results are added to the labeled samples. The updated labeled samples are then used as training samples to train the target detection network, thereby improving the detection accuracy and efficiency of the target detection network.

[0069] like Figure 3As shown, the model (i.e., the object detection network) is first trained using a small subset of labeled samples. Simultaneously, a large subset of unlabeled samples are acquired, and noise of varying intensities is added to them before being input into the object detection network. Confidence and uncertainty scores are then calculated, dividing the network into samples requiring manual annotation and relatively certain samples. The relatively certain samples are then trained iteratively, with loss curves recorded during training. These loss curves are used to filter out noisy and clean samples. Noisy samples are designated as those requiring manual annotation, while clean samples are designated as pseudo-labeled samples. Finally, the model is optimized using both pseudo-labeled and manually annotated samples to improve its accuracy.

[0070] Figure 4 A schematic block diagram of a video analysis device 400 combining active learning and noise removal, provided in an embodiment of the present invention, the device 400 comprising:

[0071] The image acquisition unit 401 is used to acquire multiple unlabeled images as initial images, and add noise of different intensities to each initial image to obtain an image sequence containing the initial image and multiple noisy images;

[0072] The target detection unit 402 is used to perform target detection on each image in the image sequence using a target detection network to obtain the target detection result for each image;

[0073] The sample screening unit 403 is used to calculate the maximum confidence score and uncertainty score of each initial image based on the target detection results, and to divide the initial image into target samples and manually labeled samples according to the calculation results;

[0074] The iterative loop unit 404 is used to perform iterative training on the labeled samples using the target detection network and record the sample loss during the iterative training process.

[0075] The video analysis unit 405 is used to select target samples with sample loss less than a preset loss threshold as pseudo-label samples, combine the pseudo-label samples and manually labeled samples to train the target detection network, and use the trained target detection network to detect and analyze the specified video images.

[0076] In one embodiment, the image acquisition unit 401 includes:

[0077] A noise addition unit is used to add noise according to the following formula to obtain the image sequence [X0, X1, X2, ... X]. i ...X n ]:

[0078] X i=X0+ε i

[0079] In the formula, noise ε i ∈N(0,σ i 2 ) follows a Gaussian distribution, σ i X represents the standard deviation, X0 represents the initial image, X i Let i represent the i-th noisy image.

[0080] In one embodiment, the target detection unit 402 includes:

[0081] The detection output unit is used to perform target detection on each image in the image sequence using the YOLOv5 target detection network, and the YOLOv5 target detection network outputs the position information b and confidence p of the target detection box of the business target in each image.

[0082] In one embodiment, the sample screening unit 403 includes:

[0083] The target sample selection unit is used to select initial images with a maximum confidence score greater than a preset confidence score and / or an uncertainty score less than a preset score threshold as target samples, and to use the remaining initial images as manually labeled samples.

[0084] In one embodiment, the sample screening unit 403 further includes:

[0085] For each initial image, the Intersection over Union (IoU) of the target detection box in the initial image and the target detection boxes in all corresponding noisy images is calculated, and the target detection box in the noisy image with the maximum IoU is selected as the matching box.

[0086] In one embodiment, the sample screening unit 403 further includes:

[0087] For each initial image, the maximum confidence α of the initial image is calculated according to the following formula:

[0088]

[0089] In the formula, p0 represents the confidence level of the target detection box in the initial image. This indicates the confidence level of the target detection box;

[0090] And the positional uncertainty β of the initial image is calculated according to the following formula:

[0091]

[0092] In the formula, b0 represents the position information of the target detection box in the initial image. This represents the position information of the matching box, where IoU represents the relationship between b0 and... The intersection and union ratio;

[0093] Traverse all noisy images, calculate the average of the maximum confidence α and the average of the location uncertainty β of the initial image, and perform a weighted sum of the average of the maximum confidence α and the average of the location uncertainty β to obtain the uncertainty score of the initial image.

[0094] In one embodiment, such as Figure 5 As shown, the iterative loop unit 404 includes:

[0095] The learning rate setting unit 501 is used to set the number of loops and set the initial learning rate for iterative training.

[0096] The decrement control unit 502 is used to control the decrease of the learning rate in each training process using an exponential decay learning rate strategy.

[0097] The learning rate recovery unit 503 is used to end the current training when the target detection network is in an overfitting state, and restore the set initial learning rate at the start of the next training session, thereby completing the iterative training.

[0098] Since the embodiments of the apparatus and the embodiments of the method correspond to each other, please refer to the description of the embodiments of the method for the embodiments of the apparatus, which will not be repeated here.

[0099] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed, can perform the steps provided in the above embodiments. The storage medium may include various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0100] This invention also provides a computer device, which may include a memory and a processor. The memory stores a computer program, and when the processor calls the computer program in the memory, it can implement the steps provided in the above embodiments. Of course, the computer device may also include various network interfaces, power supplies, and other components.

[0101] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to in the method section. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of the claims of this application.

[0102] It should also be noted that, in this specification, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

Claims

1. A video analysis method combining active learning and noise removal, characterized in that, include: Multiple unlabeled images are obtained as initial images, and noise of different intensities is added to each initial image to obtain an image sequence containing the initial images and multiple noisy images; The target detection network is used to perform target detection on each image in the image sequence to obtain the target detection result for each image; The uncertainty score of each initial image is calculated based on the target detection results, and the initial images are divided into target samples and manually labeled samples according to the calculation results; The target detection network is used to perform iterative training on the target samples, and the sample loss during the iterative training process is recorded. Target samples with a loss less than a preset loss threshold are selected as pseudo-label samples. The target detection network is trained by combining the pseudo-label samples and manually labeled samples. The trained target detection network is then used to detect and analyze the specified video images. The step of using an object detection network to perform object detection on each image in the image sequence to obtain the object detection result for each image includes: The YOLOv5 object detection network is used to perform object detection on each image in the image sequence, and the YOLOv5 object detection network outputs the location information b and confidence p of the target detection box of the business target in each image; The process of calculating the uncertainty score for each initial image based on the target detection results, and dividing the initial images into target samples and manually labeled samples according to the calculation results, includes: For each initial image, the intersection-over-union ratio (IoU) between the target detection box of the initial image and the target detection boxes of all corresponding noisy images is calculated, and the target detection box of the noisy image corresponding to the maximum IoU value is selected as the matching box. For each initial image, the maximum confidence α of the initial image is calculated according to the following formula: In the formula, This represents the confidence level of the target detection bounding box in the initial image. This indicates the confidence level of the matching box; And the positional uncertainty β of the initial image is calculated according to the following formula: In the formula, This indicates the position information of the target detection bounding box in the initial image. This represents the position information of the matching box, where IoU represents... and The intersection and union ratio; Traverse all noisy images, calculate the average of the maximum confidence α and the average of the location uncertainty β of the initial image, and perform a weighted sum of the average of the maximum confidence α and the average of the location uncertainty β to obtain the uncertainty score of the initial image.

2. The video analysis method combining active learning and noise removal according to claim 1, characterized in that, The process of acquiring multiple unlabeled images as initial images and adding noise of varying intensities to each initial image to obtain an image sequence containing the initial images and multiple noisy images includes: Noise is added according to the following formula to obtain the image sequence [X0, X1, X2, ... X]. i ...X n ]: In the formula, noise Follows a Gaussian distribution. X represents the standard deviation, X0 represents the initial image, X i Let i represent the i-th noisy image.

3. The video analysis method combining active learning and noise removal according to claim 1, characterized in that, The process of calculating the uncertainty score for each initial image based on the target detection results, and dividing the initial images into target samples and manually labeled samples according to the calculation results, includes: Initial images with uncertain scores less than a preset score threshold are selected as target samples, and the remaining initial images are used as manually labeled samples.

4. The video analysis method combining active learning and noise removal according to claim 1, characterized in that, The step of using the target detection network to perform iterative training on the target samples and recording the sample loss during the iterative training process includes: Set the number of loops and set an initial learning rate for iterative training. An exponential decay learning rate strategy is used to control the decrease of the learning rate in each training process. When the target detection network is in an overfitting state, the current training ends, and the initial learning rate is restored at the start of the next training session, thus completing the iterative training.

5. A video analysis device combining active learning and noise removal, characterized in that, include: An image acquisition unit is used to acquire multiple unlabeled images as initial images, and add noise of different intensities to each initial image to obtain an image sequence containing the initial image and multiple noisy images; The target detection unit is used to perform target detection on each image in the image sequence using a target detection network, and obtain the target detection result for each image; The step of using an object detection network to perform object detection on each image in the image sequence to obtain the object detection result for each image includes: The YOLOv5 object detection network is used to perform object detection on each image in the image sequence, and the YOLOv5 object detection network outputs the location information b and confidence p of the target detection box of the business target in each image; The sample screening unit is used to calculate the uncertainty score of each initial image based on the target detection results, and to divide the initial image into target samples and manually labeled samples according to the calculation results; The process of calculating the uncertainty score for each initial image based on the target detection results, and dividing the initial images into target samples and manually labeled samples according to the calculation results, includes: For each initial image, the intersection-over-union ratio (IoU) between the target detection box of the initial image and the target detection boxes of all corresponding noisy images is calculated, and the target detection box of the noisy image corresponding to the maximum IoU value is selected as the matching box. For each initial image, the maximum confidence α of the initial image is calculated according to the following formula: In the formula, This represents the confidence level of the target detection bounding box in the initial image. This indicates the confidence level of the matching box; And the positional uncertainty β of the initial image is calculated according to the following formula: In the formula, This indicates the position information of the target detection bounding box in the initial image. This represents the position information of the matching box, where IoU represents... and The intersection and union ratio; Traverse all noisy images, calculate the average of the maximum confidence α and the average of the location uncertainty β of the initial image, and perform a weighted sum of the average of the maximum confidence α and the average of the location uncertainty β to obtain the uncertainty score of the initial image; The iterative training unit is used to perform iterative training on the target samples using the target detection network and record the sample loss during the iterative training process. The video analysis unit is used to select target samples with sample loss less than a preset loss threshold as pseudo-label samples, combine the pseudo-label samples and manually labeled samples to train the target detection network, and use the trained target detection network to detect and analyze the specified video images.

6. A computer device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the video analysis method combining active learning and noise removal as described in any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the video analysis method combining active learning and noise removal as described in any one of claims 1 to 4.

Images