Method for generating noise label detector, noise label detection method and device

Abstract

The present disclosure provides a noise label detector generation method, a noise label detection method and device, relates to the field of computer vision, in particular to the field of image processing, the technical field of deep learning and the like. The specific implementation scheme is as follows: a first training history corresponding to any pixel in a noise sample set is obtained; wherein the noise sample set comprises a plurality of sample images, part or all of the sample images in the noise sample set comprise pixels with label errors, and the label is used to represent the category to which the pixel belongs; the first training history comprises a first prediction result of the category to which the pixel in the sample image belongs; and a noise label detector is generated by using the first training history and noise information of the noise sample set. The present disclosure can detect pixels with label errors in an image.

Description

Technical Field

[0001] This disclosure relates to the field of computer vision, and in particular to the fields of image processing, deep learning, and other technical fields. Background Technology

[0002] Image semantic segmentation is a traditional task in the field of computer vision, which aims to classify and recognize images at the pixel level.

[0003] Currently, the training performance of image segmentation models depends on a large number of images containing labeled pixels. However, obtaining a large number of images and correctly labeling each pixel in the images is extremely difficult. Using incorrectly labeled pixels to train an image segmentation model will result in low accuracy of the trained model. Summary of the Invention

[0004] This disclosure provides a method for generating a noise tag detector, a method for detecting noise tags, and an apparatus for doing so.

[0005] According to one aspect of this disclosure, a method for generating a noise label detector is provided. The noise label detector is used to detect pixels with incorrect labels in a sample image. The method includes:

[0006] Obtain the first training history corresponding to any pixel in the noisy sample set; wherein the noisy sample set includes multiple sample images, and some or all of the sample images in the noisy sample set include pixels with incorrect labels, where labels are used to characterize the category to which the pixel belongs; the first training history contains the first prediction result of the category to which the pixel in the sample image belongs; and,

[0007] A noise label detector is generated using the first training history and noise information from the noise sample set.

[0008] According to another aspect of this disclosure, a noise tag detection method is provided, comprising:

[0009] Obtain the second training history corresponding to any pixel in the image to be detected;

[0010] The image to be detected includes labeled pixels, with labels representing the category to which the pixels belong; the second training history contains a third prediction of the category to which the pixels in the image to be detected belong; and,

[0011] Based on the second training history and the noisy label detector, the predicted attributes corresponding to the pixels are obtained, and the predicted attributes corresponding to the pixels are used to characterize the accuracy of the pixel's label.

[0012] The noise tag detector is generated using any of the methods described above.

[0013] According to another aspect of this disclosure, an apparatus for generating a noise label detector is provided. The noise label detector is used to detect pixels with incorrect labels in a sample image. The apparatus includes:

[0014] The first acquisition module is used to acquire the first training history corresponding to any pixel in the noise sample set; wherein the noise sample set includes multiple sample images, and some or all of the sample images in the noise sample set include pixels with incorrect labels, where labels are used to characterize the category to which the pixel belongs; the first training history includes a first prediction result of the category to which the pixel in the sample image belongs; and...

[0015] The generation module is used to generate a noise label detector using the first training history and noise information from the noise sample set.

[0016] According to another aspect of this disclosure, a noise tag detection device is provided, comprising:

[0017] The second acquisition module is used to acquire the second training history corresponding to any pixel in the image to be detected;

[0018] The image to be detected includes labeled pixels, with labels representing the category to which the pixels belong; the second training history contains a third prediction of the category to which the pixels in the image to be detected belong; and,

[0019] The prediction module is used to obtain the predicted attributes corresponding to the pixels based on the second training history and the noisy label detector. The predicted attributes corresponding to the pixels are used to characterize the accuracy of the pixel's label.

[0020] The noise tag detector is generated by the above-mentioned device.

[0021] According to another aspect of this disclosure, an electronic device is provided, comprising:

[0022] At least one processor; and

[0023] The memory is communicatively connected to the at least one processor; wherein,

[0024] The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the methods of any embodiment of the present disclosure.

[0025] According to another aspect of this disclosure, a non-transitory computer-readable storage medium is provided storing computer instructions, wherein the computer instructions are used to cause the computer to perform a method according to any embodiment of this disclosure.

[0026] According to another aspect of this disclosure, a computer program product is provided, including a computer program that, when executed by a processor, implements a method according to any embodiment of this disclosure.

[0027] This disclosure generates a noise label detector by utilizing the first training history corresponding to any pixel in the noise sample set and the noise information of the noise sample set. This noise label detector has the characteristic of high accuracy. The noise label detector proposed in this disclosure can accurately detect pixels with incorrect labels in the image.

[0028] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0029] The accompanying drawings are provided to better understand this solution and do not constitute a limitation of this disclosure. Wherein:

[0030] Figure 1 This is a schematic diagram illustrating an application scenario according to an embodiment of this disclosure;

[0031] Figure 2 This is a flowchart illustrating a method for generating a noise tag detector according to an embodiment of the present disclosure;

[0032] Figure 3 This is a schematic block diagram of a method for generating a noise tag detector according to an embodiment of the present disclosure;

[0033] Figure 4 This is a schematic flowchart of a noise tag detection method according to an embodiment of the present disclosure;

[0034] Figure 5 This is a schematic block diagram of a noise tag detection method according to an embodiment of the present disclosure;

[0035] Figure 6 This is a schematic diagram of the structure of the noise tag detector generating apparatus 600 according to an embodiment of the present disclosure;

[0036] Figure 7 This is a schematic diagram of the structure of the noise tag detector generating apparatus 700 according to an embodiment of the present disclosure;

[0037] Figure 8 This is a schematic diagram of the structure of a noise tag detection device 800 according to an embodiment of the present disclosure;

[0038] Figure 9 A schematic block diagram of an example electronic device 900 that can be used to implement embodiments of the present disclosure is shown. Detailed Implementation

[0039] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0040] In this document, the term "and / or" merely describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. The term "at least one" in this document indicates any combination of at least two of a plurality of elements. For example, including at least one of A, B, and C can mean including any one or more elements selected from the set consisting of A, B, and C. The terms "first" and "second" in this document refer to and distinguish between multiple similar technical terms, not to restrict the order or to limit there to only two. For example, "first feature" and "second feature" refer to two categories / two features; the first feature can be one or more, and the second feature can also be one or more.

[0041] Image semantic segmentation is a traditional task in computer vision, aiming to classify and recognize images at the pixel level. Currently, the training effect of image segmentation models depends on a large number of images containing labeled pixels; however, obtaining a large number of images and the correct labels for each pixel in the images is extremely difficult.

[0042] Generally, training an image segmentation model using incorrectly labeled pixels can negatively impact the accuracy of the trained model. Therefore, to improve the accuracy of the image segmentation model from the perspective of the training set, this disclosure proposes a method for generating a noise label detector. This noise label detector can detect incorrectly labeled pixels in sample images, minimizing the negative impact of incorrectly labeled pixels on the accuracy of the image segmentation model during training.

[0043] Figure 1 This is a schematic diagram illustrating an application scenario according to an embodiment of this disclosure. For example... Figure 1As shown, the application scenario of the noise label detector proposed in this embodiment can include a terminal device 110 and a server 120. Specifically, the terminal device 110 can send an image to be detected to the server 120, and the server 120 can use a noise label detector (which can be obtained using the noise label detector generation method proposed in this embodiment) to detect the image to obtain a detection result, wherein the detection result can include pixels with incorrect labels in the image to be detected. Figure 1 As shown, server 120 can also send detection results back to terminal device 110. It should be noted that this embodiment does not limit the number of terminal devices 110. Figure 1 For example, the application scenario of this noise tag detector may include one or more terminal devices 110, such as smartphones, tablets, and laptops. Of course, this disclosure does not limit the type of terminal device; the above is merely an example. For instance, the terminal device 110 proposed in this disclosure may also include desktop computers, wearable devices, smart home devices, and head-mounted devices, etc. In some embodiments, the terminal device 110 proposed in this disclosure can interact with the server 120 via wireless networks, wired networks, and removable storage media. It should be noted that the server 120 may include independent physical servers, server clusters composed of multiple physical servers, distributed systems, and cloud servers providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), big data, and artificial intelligence platforms. Furthermore, the application scenario proposed in this disclosure may also include the terminal device 110 alone. Specifically, the noise tag detector can be directly deployed in the terminal device 110. At this time, terminal device 110 may include a single server, a server cluster consisting of several servers, and a cloud computing service center.

[0044] Figure 2This is a flowchart illustrating a method for generating a noise label detector according to an embodiment of the present disclosure. The noise label detector is used to detect pixels with incorrect labels in a sample image. This method can be applied to a noise label detector generation apparatus. For example, the apparatus can be deployed in a single-machine, multi-machine, or cluster system on a terminal, server, or other processing device. The terminal can be a user equipment (UE), mobile device, personal digital assistant (PDA) handheld device, computing device, in-vehicle device, wearable device, etc. In some possible implementations, the method can also be implemented by a processor calling computer-readable instructions stored in memory. Figure 2 As shown, the method for generating the noise tag detector includes:

[0045] S210. Obtain the first training history corresponding to any pixel in the noise sample set; wherein, the noise sample set includes multiple sample images, and some or all of the sample images in the noise sample set include pixels with incorrect labels, the label being used to characterize the category to which the pixel belongs; the first training history includes a first prediction result of the category to which the pixel in the sample image belongs; and,

[0046] S220. Using the first training history and the noise information of the noise sample set, generate a noise label detector.

[0047] The first training history may include the training trace corresponding to any pixel in the noisy sample set. The first training history may be represented by time series data. The first training history provides the features needed to distinguish pixels with incorrect labels from pixels with correct labels.

[0048] In one example, some or all of the sample images in the noisy sample set may include pixels with incorrect labels, meaning that the category represented by the label corresponding to the pixel is different from the actual category to which the pixel belongs. For example, assuming the sample image is a face image, if the pixel categories in the face image include background, face, eyes, and mouth, then if the label representing the category of any pixel in the face image is background, but the actual category to which the pixel belongs is face, then that pixel is determined to be a pixel with an incorrect label.

[0049] Of course, some or all of the sample images in the noise sample set proposed in this disclosure embodiment may also include pixels with blurred labels, that is, the label corresponding to the pixel cannot clearly represent the category to which the pixel belongs, or the category represented by the label corresponding to the pixel is ambiguous. For example, taking the sample image as a face image, and the categories to which the pixels in the face image belong include background, face, eyes, and mouth, if the label of any pixel in the face image is blank, it is determined that the label corresponding to the pixel cannot clearly represent the category to which the pixel belongs; or, if the category represented by the label of any pixel in the face image includes background and face, it is determined that the category represented by the label corresponding to the pixel is ambiguous.

[0050] The noise label detector generation method proposed in this disclosure can reduce the training time and computational cost required to generate the noise label detector while ensuring its accuracy. Furthermore, the noise label detector generated using this disclosure can accurately detect pixels with incorrect labels in a sample image based on the first training history corresponding to pixels in the sample image.

[0051] Accordingly, some or all of the sample images in the noise sample set used in the embodiments of this disclosure may refer to randomly obtained images, which include at least one pixel with an incorrect label.

[0052] Here, since it cannot be guaranteed that a large number of images containing mislabeled pixels can be collected under random conditions, or that the randomly acquired images contain a sufficient number of mislabeled pixels, the noisy sample set composed of randomly acquired images has at least the following problems: the number of images is too small, and the number of mislabeled pixels in the images is too small. Therefore, the noisy sample set composed of randomly acquired images (i.e., a set including multiple randomly acquired images) cannot obtain a first training history that meets the requirements.

[0053] Accordingly, in order to solve the above problems, this disclosure proposes a method for obtaining a noise sample set. The method can perform noise processing on the original sample set to obtain a noise sample set. Some or all of the sample images in the original sample set may not include pixels with incorrect labels, or may include fewer pixels with incorrect labels.

[0054] However, obtaining sample images that do not contain incorrect labels (i.e., all pixels in the sample image are correctly labeled) is difficult, costly, and time-consuming. Therefore, it is generally acceptable to use an original sample set consisting of sample images containing fewer incorrect labels (i.e., the sample image contains fewer pixels with incorrect labels). For example, an original sample set consisting of sample images containing fewer incorrect labels could be an object detection dataset (Pascal VOC) or a deep learning image dataset (MS COCO). Here, the specific method for using an original sample set consisting of sample images containing fewer incorrect labels includes at least the following:

[0055] The original sample set is processed to remove noise, resulting in the noisy sample set.

[0056] Where the number of pixels with incorrect labels in the original sample set is less than or equal to the first threshold, and the number of pixels with incorrect labels in the noisy sample set is greater than or equal to the second threshold;

[0057] The second threshold is greater than or equal to the first threshold.

[0058] In some implementations, if D is used to represent the original sample set, D N Let D represent the noise sample set. Then, compared to D... N D contains fewer pixels with incorrect labels.

[0059] Using the original sample set to obtain the noise sample set can ensure that the noise sample set contains a sufficient number of pixels with incorrect labels, and at the same time improve the robustness of the pixels with incorrect labels in the noise sample set, thus providing a certain guarantee for the subsequent generation of a noise label detector with higher accuracy.

[0060] Figure 3 This is a schematic block diagram of a method for generating a noise tag detector according to an embodiment of the present disclosure. Figure 3 As shown, after obtaining the noise sample set D N Then, the noise sample set D can be used. N The deep model is trained to obtain the noise sample set D. N The first training history corresponding to any pixel in the dataset.

[0061] Specifically, obtaining the first training history corresponding to any pixel in the noise sample set includes:

[0062] The deep model is trained using this noise sample set to obtain the first training history corresponding to any pixel in the noise sample set.

[0063] This deep model can include models such as fully convolutional networks (FCN).

[0064] In addition, the first training history may include the training history obtained by any pixel in the noise sample set in one training session; or, the first training history may include the training history obtained by any pixel in the noise sample set in multiple training sessions (e.g., t times, where t is an integer greater than 1), that is, the set of t training histories obtained by any pixel in the noise sample set in t training sessions.

[0065] Accordingly, the noise sample set D at the t-th training time can be determined using equation (1). N The first training history corresponding to pixel (h, w) in the i-th sample image:

[0066]

[0067] Among them, L N c (i, h, w, t) represents the noise sample set D during the t-th training iteration. N The first training history (i.e., the representation of the noisy sample set D) corresponding to pixel (h, w) in the i-th sample image. N The prediction result when the pixel (h, w) in the i-th sample image belongs to class c); x i，h，w Represents the noise sample set D N The pixel (h, w) on the i-th sample image; y N i，h，w Represents the noise sample set D N The label corresponding to pixel (h, w) in the i-th sample image; f represents the depth model; Represents the noise sample set D N The pixels (h, w) on the i-th sample image are trained on the deep model f for the t-th time.

[0068] It should be noted that, in this embodiment of the present disclosure, before training the deep model for the tth time using the noisy sample set, the parameters of the deep model can be adjusted using the training history obtained during the (t-1)th training session, thereby improving the accuracy of the first training history output by the deep model. For example, the parameters of the deep learning model can be adjusted based on the backpropagation (BP) algorithm, the stochastic gradient descent (SGD) algorithm, or a variant thereof.

[0069] This embodiment of the invention, while ensuring the accuracy of the first training history, can reduce the time consumed in acquiring the first training history and improve the efficiency of acquiring the first training history. Simultaneously, because the method for acquiring the first training history proposed in this embodiment reduces the error of the first training history, it can improve the accuracy of the subsequently generated noise label detector.

[0070] like Figure 3 As shown, the noise label detector generation method proposed in this embodiment of the present disclosure also requires prior acquisition of noise information in the noise sample set, so as to subsequently use the first training history and the noise information in the noise sample set to generate the noise label detector.

[0071] Here, the noise information of the noise sample set can be used to characterize the first true attribute corresponding to any pixel in the noise sample set, which includes whether the label is incorrect or correct.

[0072] For example, taking K as the representation of noise information in a noise sample set, the noise information K in the noise sample set can include multiple k. i，h，w Specifically, the k i，h，w It can represent the noise sample set D N The first true attribute corresponding to pixel (h, w) in the i-th sample image. For example, when k i，h，w When k = 1, it indicates that the first true attribute of pixel (h, w) is a label error, meaning that the label corresponding to pixel (h, w) in the noise sample set does not match the true label of that pixel; when k i，h，w A value of 0 indicates that the first true attribute of the pixel is correctly labeled, meaning that the label corresponding to pixel (h, w) in the noise sample set matches the true label of that pixel. That is, the noise sample set D... N The first true attribute corresponding to the pixel in the i-th sample image is shown in Equation (2):

[0073]

[0074] Among them, y i，h，w N represents the noise sample set D. N The actual label corresponding to the middle pixel (h, w); y i，h，w This indicates the correct label corresponding to pixel (h, w).

[0075] Of course, the embodiments disclosed herein do not limit k i，h，w For example, it can also be represented in k. i，h，w When k = 1, it means that the first true attribute of the pixel (h, w) is correctly labeled; i，h，w When = 0, it indicates that the first true attribute of the pixel (h, w) is a label error.

[0076] The embodiments of this disclosure characterize the first true attribute corresponding to any pixel in the noise sample set by using the noise information of the noise sample set, thereby improving the accuracy of determining the first true attribute corresponding to any pixel in the noise sample set.

[0077] The above briefly introduces how to obtain a noise sample set, the first training history corresponding to any pixel in the noise sample set, and how to obtain noise information in the noise sample set.

[0078] like Figure 3 As shown in the embodiments of this disclosure, the method for generating a noise label detector can generate a noise label detector by using the first training history and the noise information of the noise sample set after obtaining the noise sample set, the first training history, and the noise information of the noise sample set. The following will provide a detailed explanation of how to generate the noise label detector.

[0079] In this embodiment of the disclosure, the noise information of the noise sample set can be used to obtain the first real attribute corresponding to any pixel in the sample image, and the noise label detector can be generated based on the first training history corresponding to the pixel and the first real attribute corresponding to the pixel.

[0080] At this point, generating a noise label detector using the first training history and the noise information from the noise sample set can include:

[0081] Input the first training history corresponding to any pixel into the initial noise label detector to obtain the predicted attribute corresponding to the pixel; and determine the first true attribute corresponding to the pixel based on the noise information of the noise sample set.

[0082] Using the predicted attribute corresponding to the pixel and the first true attribute, calculate the loss function of the initial noise label detector; and,

[0083] The initial noise label detector is adjusted according to the loss function to obtain the final noise label detector.

[0084] In one example, when training a noise label detector, the [first training history of a pixel] can be input into the noise label detector, and the noise label detector can output the predicted attributes of that pixel.

[0085] It should be noted that the predicted attribute corresponding to this pixel can be used to characterize the accuracy of the label for that pixel. The accuracy of the label for that pixel, as characterized by the predicted attribute, can be determined by the noise information of the noisy sample set.

[0086] Taking any rational number in the range [0, 1] as an example, if k is a rational number in the noise information K of the noise sample set, then...i，h，w =1 indicates that the first true attribute of pixel (h, w) is a label error, k i，h，w When k = 0, it indicates that the first true attribute of the pixel is correctly labeled. In this case, if the predicted attribute output by the initial noise label detector for any pixel is 0.9, it can be considered that the initial noise label detector's prediction result for that pixel is 90% likely to be incorrectly labeled. If the predicted attribute output by the initial noise label detector for any pixel is 0.1, it can be considered that the initial noise label detector's prediction result for that pixel is 90% likely to be correctly labeled, which is equivalent to a 10% probability of being incorrectly labeled. Conversely, if in the noise information K of the noise sample set, k... i，h，w When k = 0, it indicates that the first true attribute of pixel (h, w) is a label error. i，h，w When the value is 1, it indicates that the first true attribute of the pixel is that the label is correct. In this case, if the prediction attribute output by the initial noise label detector for any pixel is 0.9, it can be considered that the prediction result of the initial noise label detector for that pixel is 90% likely to be correctly labeled. If the prediction attribute output by the initial noise label detector for any pixel is 0.1, it can be considered that the prediction result of the initial noise label detector for that pixel is 10% likely to be correctly labeled, which means that the probability of the pixel being incorrectly labeled is 90%.

[0087] In some implementations, k i，h，w =1 indicates that the first true attribute of pixel (h, w) is a label error, k i，h，w =0 indicates that the first true attribute of the pixel is correctly labeled. If the predicted attribute output by the noise label detector is 0.1, while the first true attribute corresponding to the pixel is 1, it means that the noise label detector cannot detect samples with incorrect labels. Therefore, the noise label detector has a large error. In this case, the loss function can be determined based on the predicted attribute output by the initial noise label detector and the first true attribute of the pixel.

[0088] If the predicted attribute output by the noise label detector is 0.9 and the first true attribute corresponding to the pixel is 1, then the result indicates that the noise label detector can detect samples with incorrect labels, and therefore the noise label detector does not have a large error.

[0089] Accordingly, using the predicted attribute corresponding to the pixel and the first true attribute, the loss function of the initial noise label detector is calculated, including:

[0090] The first training history corresponding to the pixels in the noise sample set is input into the initial noise label detector to obtain the predicted attribute corresponding to the pixel in the sample image; and the binary cross-entropy loss function is calculated based on the first true attribute corresponding to the pixel in the sample image and the predicted attribute of the pixel; the binary cross-entropy loss functions corresponding to each pixel in the sample image are added together to obtain the loss function corresponding to the sample image.

[0091] Furthermore, this loss function can be used to adjust the initial noise tag detector to obtain the noise tag detector. That is, the noise tag detector is determined using equation (3):

[0092]

[0093] in, represents the trained noise label detector, and g represents the initial noise label detector; Represents the noise sample set D N The first training history corresponding to pixel (h, w) in the i-th sample image. Represents the noise sample set D N The predicted attribute corresponding to pixel (h, w) in the i-th sample image; k i，h，w Represents the noise sample set D N The first true attribute corresponding to pixel (h, w) in the i-th sample image; This represents the binary cross-entropy loss function used to calculate a and b.

[0094] The initial noise label detector can include a Long Short-Term Memory (InceptionTime) model, which excels at handling temporal information. This InceptionTime model remains effective even for sets containing only correctly labeled pixels. Furthermore, during the generation of the noise label detector, other category information can be incorporated to significantly increase the number of parameters and computational cost, thereby improving the robustness of the noise label detector.

[0095] It should be noted that the embodiments of this disclosure may employ an optimizer to adjust the initial noise tag detector, wherein the optimizer may include AdamW.

[0096] The noise label detector generated by utilizing the first training history corresponding to any pixel in the sample image and the first true attribute corresponding to that pixel not only has the advantage of high accuracy, but also reduces the time required to generate the noise label detector.

[0097] However, since some or all of the sample images in the noise sample set include pixels with incorrect labels, the number of pixels with incorrect labels in the noise sample set is relatively large. If the first training history corresponding to all the pixels with incorrect labels in the noise sample set and the noise information of the noise sample set are used to train the noise label detector, it will consume a lot of unnecessary time and cost.

[0098] Therefore, to address the aforementioned problems, this disclosure proposes a method for generating a noise label detector by utilizing the first training history corresponding to any region in a sample image and the noise information corresponding to that region. This method can reduce the cost and time required to generate the noise label detector while ensuring its accuracy.

[0099] Specifically, using the first training history and the noise information of the noise sample set, a noise label detector is generated, including:

[0100] The sample image is divided into multiple regions, and each region contains multiple pixels.

[0101] For any given region, determine the loss function of the initial noise label detector; and adjust the initial noise label detector according to the loss function to obtain the final noise label detector.

[0102] The method for generating a noise label detector for any region in a sample image proposed in this disclosure can reduce the computational and storage costs required to generate the noise label detector while ensuring its accuracy.

[0103] The sample image is divided into multiple regions, including:

[0104] The sample image is divided into multiple regions based on at least one of the semantic features, texture features, color features, and brightness features of multiple pixels in the sample image; and / or,

[0105] The sample image is divided into multiple regions according to a preset size.

[0106] For example, the sample image can be divided into multiple superpixel regions based on at least one of the semantic features, texture features, color features, and brightness features of multiple pixels in the sample image. A superpixel region can be a small region composed of a series of adjacent pixels with similar features such as color, brightness, and texture. These small regions mostly retain the effective information of the image and generally do not destroy the boundary information contained in the image. Specifically, the sample image can be segmented into multiple superpixel regions using the Simple Linear Iterative Clustering (SLIC) algorithm or the Quickshift image segmentation algorithm.

[0107] Of course, the embodiments of this disclosure can also divide the sample image according to a preset size. For example, to achieve the required accuracy for the label noise detector, the sample image can be divided into multiple regions of 10x10 pixels each. It should be noted that the present disclosure can also divide the sample image into multiple regions of different sizes; for example, the sample image can be divided into one 10x10 pixel region and two 20x20 pixel regions. Of course, the embodiments of this disclosure do not limit the size of the region; the above are merely examples.

[0108] The sample image segmentation method proposed in this embodiment not only does not destroy the boundary information contained in the image, but also ensures that the segmented regions can retain the effective information in the sample image, thus avoiding the existence of regions that do not contain effective information.

[0109] Accordingly, this embodiment of the disclosure also requires determining the loss function of the initial noise tag detector for any region, and using the loss function to adjust the initial noise tag detector to obtain the noise tag detector.

[0110] Specifically, for any given region, the loss function for the initial noise label detector is determined, including:

[0111] Based on the first training history corresponding to the pixels in the region, calculate the second prediction result of the category to which the region belongs;

[0112] The second predicted category of the region is input into the initial noise label detector to obtain the predicted attribute corresponding to the region; and based on the noise information of the noise sample set, the second true attribute corresponding to the region is determined; and,

[0113] Using the predicted attribute corresponding to the region and the second true attribute, the loss function of the initial noise label detector is calculated.

[0114] It should be noted that the predicted attributes corresponding to this region can be used to characterize the accuracy of the labels in that region. These predicted attributes can include the proportion of pixels with incorrect labels in the region, or the proportion of pixels with correct labels in the region.

[0115] Accordingly, using the predicted attribute corresponding to the region and the second true attribute, the loss function of the initial noise label detector is calculated, including:

[0116] Based on the first training history corresponding to multiple pixels in the region, the second detection result corresponding to the region is obtained, and the second detection result is input into the initial noise label detector to obtain the predicted attribute corresponding to the region in the sample image; and based on the second true attribute corresponding to the region in the sample image and the predicted attribute of the pixel, the binary cross-entropy loss function is calculated; the binary cross-entropy loss functions corresponding to each region in the sample image are added together to obtain the loss function corresponding to the sample image.

[0117] Furthermore, the initial noise tag detector can be adjusted using this loss function to obtain the noise tag detector. That is, the noise tag detector is determined using equation (4):

[0118]

[0119] in, represents the noise tag detector, and g represents the initial noise tag detector; Represents the noise sample set D N The second detection result corresponding to region m in the i-th sample image. Represents the noise sample set D N The predicted attribute corresponding to region m on the i-th sample image; k i，m Represents the noise sample set D N The second true attribute corresponding to region m in the i-th sample image; This represents the binary cross-entropy loss function used to calculate a and b.

[0120] The loss function obtained by using the embodiments of this disclosure can quickly converge the initial noise label detector, thereby reducing the time required to generate the noise label detector while ensuring the accuracy of the noise label detector.

[0121] Furthermore, based on the noise information of the noise sample set, the second true attribute corresponding to the region is determined, including:

[0122] Using the noise information from this noise sample set, determine the pixels with incorrect labels and / or pixels with correct labels in this region;

[0123] The second true attribute of a region is determined based on the number of pixels with incorrect labels and / or the number of pixels with correct labels, as well as the number of pixels contained in the region.

[0124] The noise label detector generated by utilizing the second detection result corresponding to any region in the sample image and the second true attribute corresponding to that region not only has the advantage of high accuracy, but also reduces the storage and computation costs required to generate the noise label detector.

[0125] The second real attribute can be used to characterize at least one of the following:

[0126] The proportion of pixels with incorrect labels in this area;

[0127] The proportion of pixels with correct labels in this area.

[0128] In one example, when training the noise label detector, the second detection result of the region can be input into the initial noise label detector, which then outputs the predicted attributes of the region.

[0129] Taking the example where both the second true attribute and the predicted attribute are used to characterize the proportion of pixels with incorrect labels in a region. If both the predicted attribute and the second true attribute are any rational numbers in the range [0,1], if the noise label detector detects that the predicted attribute of the region is 0, then it is considered that the prediction result of the noise label detector for the region is: the proportion of pixels with incorrect labels in the region is 0; if the noise label detector detects that the predicted attribute of the region is 1, then it is considered that the prediction result of the noise label detector for the region is: the proportion of pixels with incorrect labels in the region is 1.

[0130] For example, if the predicted attribute output by the noise label detector is 0.6, while the true attribute of the region is 0.4, it proves that the proportion of pixels with incorrect predicted labels in the region is not accurate enough. In this case, the loss function can be determined based on the predicted attribute output by the noise label detector and the second true attribute of the region.

[0131] The present invention proposes to improve the accuracy of determining the second true attribute corresponding to the region by utilizing the proportion of pixels with incorrect labels or the proportion of pixels with correct labels in the region, thereby improving the accuracy of subsequent noise label detection.

[0132] This disclosure also proposes a method for detecting noise tags. Figure 4 This is a flowchart illustrating a noise tag detection method according to an embodiment of the present disclosure. Noise tag detection can be performed using any of the noise tag detectors proposed in the above embodiments, such as... Figure 4 As shown, the noise tag detection method includes:

[0133] S410. Obtain the second training history corresponding to any pixel in the image to be detected; wherein, the image to be detected includes labeled pixels, the labels being used to characterize the category to which the pixels belong; the second training history includes a third prediction result of the category to which the pixels in the image to be detected belong; and,

[0134] S420. Based on the second training history and the noise label detector, obtain the predicted attributes corresponding to the pixels. The predicted attributes corresponding to the pixels are used to characterize the accuracy of the pixel's label.

[0135] The noise tag detector is generated using any of the noise tag detector generation methods.

[0136] The noise label detection method proposed in this disclosure can quickly and efficiently detect pixels with incorrect labels in the image to be detected based on the second training history corresponding to the pixels in the image to be detected.

[0137] Figure 5 This is a schematic block diagram of a noise tag detection method according to an embodiment of the present disclosure. Figure 5 As shown, a deep model can be trained using the image to be detected to obtain a second training history corresponding to any pixel in the image. This second training history is then input into a noise label detector to obtain pixels with incorrect labels in the image. Figure 5 As shown, the noise tag detector can be obtained using the noise tag detector generation method described above.

[0138] In some implementations, the predicted attributes are obtained based on the second training history and the noise label detector, including:

[0139] The image to be detected is divided into multiple regions to be detected, and each region to be detected includes multiple pixels.

[0140] Based on the second training history corresponding to the pixels in any region to be detected, calculate the fourth prediction result of the category to which the region to be detected belongs;

[0141] The fourth prediction result of the category to which the region to be detected belongs is input into the noise label detector to obtain the prediction attribute corresponding to the region to be detected. The prediction attribute corresponding to the region to be detected is used to characterize the accuracy of the label of the pixel in the region to be detected.

[0142] The method proposed in this disclosure for detecting noise labels in any region of an image does not require processing every single pixel, which can greatly reduce computation and storage costs. Therefore, it can reduce the cost of noise label detection while ensuring the accuracy of the noise label detection results.

[0143] A prediction attribute is set to characterize the proportion of pixels with incorrect labels in the region. Taking any rational number in the range [0,1] as an example, if the noise label detector detects a prediction attribute of 0.3 for the region, it is considered that the prediction result of the noise label detector for the region is: the proportion of pixels with incorrect labels in the region is 30%; if the noise label detector detects a prediction attribute of 0.9 for the region, it is considered that the prediction result of the noise label detector for the region is: the proportion of pixels with incorrect labels in the region is 90%.

[0144] In some implementations, the image to be detected is divided into multiple regions to be detected, including:

[0145] The image to be detected is divided into multiple regions to be detected based on at least one of the semantic features, texture features, color features, and brightness features of multiple pixels in the image; and / or,

[0146] The image to be detected is divided into multiple regions according to a preset size.

[0147] The image segmentation method proposed in this embodiment not only does not destroy the boundary information contained in the image, but also ensures that the segmented regions can retain the effective information of the image to be detected, thereby improving the efficiency of label noise detection.

[0148] In some implementations, the predicted attributes are obtained based on the second training history and the noise label detector, including:

[0149] The second training history corresponding to any pixel in the image to be detected is input into the noise label detector to obtain the predicted attribute corresponding to the pixel. The predicted attribute corresponding to the pixel is used to characterize the accuracy of the label of the pixel.

[0150] When the prediction attribute is set to 1, it indicates that the pixel is incorrectly labeled; when the prediction attribute is 0, it indicates that the pixel is correctly labeled. If the initial noise label detector detects a pixel with a prediction attribute of 0.9, it is considered that the initial noise label detector's prediction result for that pixel is 90% likely to be incorrectly labeled. If the initial noise label detector predicts a pixel with a prediction attribute of 0.1, it is considered that the initial noise label detector's prediction result for that pixel is 10% likely to be incorrectly labeled, which is equivalent to a 90% likelihood that the pixel is correctly labeled.

[0151] By using the second training history corresponding to any pixel in the image to be detected and the noise label detector, the predicted attribute corresponding to that pixel has high accuracy and the time required to obtain the predicted attribute corresponding to that pixel is short.

[0152] In some implementations, obtaining the second training history corresponding to any pixel in the image to be detected includes:

[0153] The deep model is trained using the image to be detected in order to obtain the second training history corresponding to any pixel in the image to be detected.

[0154] The method for obtaining the second training history proposed in this disclosure can reduce the time consumed in obtaining the second training history while ensuring its accuracy.

[0155] This disclosure also proposes an apparatus for generating a noise label detector, which is used to detect pixels with incorrect labels in a sample image. Figure 6 This is a schematic diagram of the structure of a noise tag detector generation apparatus 600 according to an embodiment of the present disclosure, including:

[0156] The first acquisition module 610 is used to acquire the first training history corresponding to any pixel in the noise sample set; wherein the noise sample set includes multiple sample images, and some or all of the sample images in the noise sample set include pixels with incorrect labels, where labels are used to characterize the category to which the pixels belong; the first training history includes a first prediction result of the category to which the pixels in the sample images belong; and

[0157] The generation module 620 is used to generate a noise label detector using the first training history and noise information from the noise sample set.

[0158] In some implementations, noise information from the noise sample set is used to characterize a first true attribute corresponding to any pixel in the noise sample set, the first true attribute including whether the label is incorrect or correct.

[0159] Figure 7This is a schematic diagram of the structure of the noise tag detector generation apparatus 700 according to an embodiment of the present disclosure, as shown below. Figure 7 As shown, in some embodiments, the generation module 620 further includes:

[0160] The segmentation submodule 621 is used to divide the sample image into multiple regions, each region comprising multiple pixels;

[0161] The determination submodule 622 is used to determine the loss function of the initial noise tag detector for any region; and to adjust the initial noise tag detector according to the loss function to obtain the final noise tag detector.

[0162] In some implementations, the determining submodule 622 is used for:

[0163] Based on the first training history corresponding to the pixels in the region, calculate the second prediction result of the category to which the region belongs;

[0164] The second predicted category of the region is input into the initial noise label detector to obtain the predicted attribute corresponding to the region; and based on the noise information of the noise sample set, the second true attribute corresponding to the region is determined; and,

[0165] Using the predicted attribute corresponding to the region and the second true attribute, the loss function of the initial noise label detector is calculated.

[0166] In some implementations, the determining submodule 622 is used for:

[0167] Using the noise information from this noise sample set, determine the pixels with incorrect labels and / or pixels with correct labels in this region;

[0168] The second true attribute of a region is determined based on the number of pixels with incorrect labels and / or the number of pixels with correct labels in the region, as well as the number of pixels contained in the region.

[0169] In some implementations, the second real property is used to characterize at least one of the following:

[0170] The proportion of pixels with incorrect labels in this area;

[0171] The proportion of pixels with correct labels in this area.

[0172] In some implementations, the partitioning submodule 621 is used for:

[0173] The sample image is divided into multiple regions based on at least one of the semantic features, texture features, color features, and brightness features of multiple pixels in the sample image; and / or,

[0174] The sample image is divided into multiple regions according to a preset size.

[0175] In some implementations, the generation module 620 is used for:

[0176] Input the first training history corresponding to any pixel into the initial noise label detector to obtain the predicted attribute corresponding to the pixel; and determine the first true attribute corresponding to the pixel based on the noise information of the noise sample set.

[0177] Using the predicted attribute corresponding to the pixel and the first true attribute, calculate the loss function of the initial noise label detector; and,

[0178] The initial noise label detector is adjusted according to the loss function to obtain the final noise label detector.

[0179] In some implementations, the first acquisition module 610 is used to train the depth model using the noise sample set to obtain the first training history corresponding to any pixel in the noise sample set.

[0180] In some implementations, it also includes:

[0181] Processing module 730 is used to process the noise in the original sample set to obtain a noisy sample set;

[0182] Among them, the number of pixels with incorrect labels in the original sample set is less than or equal to the first threshold, and the number of pixels with incorrect labels in the noisy sample set is greater than or equal to the second threshold.

[0183] The second threshold is greater than or equal to the first threshold.

[0184] This disclosure also proposes a noise tag detection device. Figure 8 This is a schematic diagram of the structure of a noise tag detection device 800 according to an embodiment of the present disclosure, including:

[0185] The second acquisition module 810 is used to acquire the second training history corresponding to any pixel in the image to be detected;

[0186] The image to be detected includes labeled pixels, which represent the category to which the pixel belongs; the second training history contains a third prediction result of the category to which the pixels in the image to be detected belong; and,

[0187] The prediction module 820 is used to obtain the predicted attribute corresponding to the pixel based on the second training history and the noise label detector. The predicted attribute corresponding to the pixel is used to characterize the accuracy of the label of the pixel. The noise label detector is generated by any of the noise label detector generation devices.

[0188] In some implementations, the prediction module 820 is used for:

[0189] The image to be detected is divided into multiple regions to be detected, and each region to be detected includes multiple pixels.

[0190] Based on the second training history corresponding to the pixels in any region to be detected, calculate the fourth prediction result of the category to which the region to be detected belongs;

[0191] The fourth prediction result of the category to which the region to be detected belongs is input into the noise label detector to obtain the prediction attribute corresponding to the region to be detected. The prediction attribute corresponding to the region to be detected is used to characterize the accuracy of the label of the pixel in the region to be detected.

[0192] In some implementations, the prediction module 820 is used for:

[0193] The image to be detected is divided into multiple regions to be detected based on at least one of the semantic features, texture features, color features, and brightness features of multiple pixels in the image; and / or,

[0194] The image to be detected is divided into multiple regions according to a preset size.

[0195] In some implementations, the prediction module 820 is used to input the second training history corresponding to any pixel in the image to be detected into the noise label detector to obtain the prediction attribute corresponding to the pixel, which is used to characterize the accuracy of the label of the pixel.

[0196] In some implementations, the second acquisition module 810 is used to train a deep model using the image to be detected in order to obtain a second training history corresponding to any pixel in the image to be detected.

[0197] The specific functions and examples of each module and submodule of the apparatus in this disclosure can be found in the relevant descriptions of the corresponding steps in the above method embodiments, and will not be repeated here.

[0198] The acquisition, storage, and application of user personal information involved in the technical solution disclosed herein comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0199] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0200] Figure 9A schematic block diagram of an example electronic device 900 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0201] like Figure 9 As shown, device 900 includes a computing unit 901, which can perform various appropriate actions and processes based on a computer program stored in read-only memory (ROM) 902 or a computer program loaded from storage unit 908 into random access memory (RAM) 903. RAM 903 may also store various programs and data required for the operation of device 900. The computing unit 901, ROM 902, and RAM 903 are interconnected via bus 904. Input / output (I / O) interface 905 is also connected to bus 904.

[0202] Multiple components in device 900 are connected to I / O interface 905, including: input unit 906, such as keyboard, mouse, etc.; output unit 907, such as various types of monitors, speakers, etc.; storage unit 908, such as disk, optical disk, etc.; and communication unit 909, such as network card, modem, wireless transceiver, etc. Communication unit 909 allows device 900 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0203] The computing unit 901 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 901 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 901 performs the various methods and processes described above, such as the noise tag detector generation method and the noise tag detection method. For example, in some embodiments, the noise tag detector generation method and the noise tag detection method can be implemented as computer software programs tangibly contained in a machine-readable medium, such as storage unit 908. In some embodiments, part or all of the computer program can be loaded and / or installed on device 900 via ROM 902 and / or communication unit 909. When the computer program is loaded into RAM 903 and executed by the computing unit 901, one or more steps of the noise tag detector generation method and the noise tag detection method described above can be performed. Alternatively, in other embodiments, the computing unit 901 may be configured by any other suitable means (e.g., by means of firmware) to perform the noise tag detector generation method and the noise tag detection method.

[0204] Various embodiments 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-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0205] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0206] In the context of this disclosure, a machine-readable medium can 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 can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0207] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0208] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with embodiments of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0209] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0210] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0211] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for generating a noise label detector, wherein the noise label detector is used to detect pixels with incorrect labels in a sample image, the method comprising: Obtain the first training history corresponding to any pixel in the noise sample set; wherein, the noise sample set includes multiple sample images, and some or all of the sample images in the noise sample set include pixels with incorrect labels, the labels being used to characterize the category to which the pixel belongs; the first training history includes a first prediction result of the category to which the pixel in the sample image belongs; and, A noise label detector is generated using the first training history and the noise information of the noise sample set; The noise label detector is generated by: dividing the sample image into multiple regions, each region comprising multiple pixels; each region including a superpixel region, which is composed of pixels that are adjacent in position and have similar color, brightness, and texture; calculating a second prediction result of the category to which the region belongs based on a first training history corresponding to the pixels in the region; inputting the second prediction result into an initial noise label detector to obtain a predicted attribute corresponding to the region; determining a second true attribute corresponding to the region based on the noise information of the noise sample set; calculating a binary cross-entropy loss function corresponding to the region using the predicted attribute and the second true attribute; summing the binary cross-entropy loss functions corresponding to each region to obtain the loss function of the initial noise label detector; the predicted attribute including the proportion of incorrectly labeled pixels in the region, or the proportion of correctly labeled pixels in the region; and adjusting the initial noise label detector according to the loss function to obtain the final noise label detector.

2. The method according to claim 1, wherein, The noise information of the noise sample set is used to characterize the first true attribute corresponding to any pixel in the noise sample set, and the first true attribute includes whether the label is incorrect or the label is correct.

3. The method according to claim 1, wherein, Determining the second true attribute corresponding to the region based on the noise information of the noise sample set includes: Using the noise information from the noise sample set, determine the pixels in the region with incorrect labels and / or pixels with correct labels; The second true attribute corresponding to the region is determined based on the number of pixels with incorrect labels and / or the number of pixels with correct labels in the region, as well as the number of pixels contained in the region.

4. The method according to claim 1 or 3, wherein the second true attribute is used to characterize at least one of the following: The proportion of pixels with incorrect labels in the region; The proportion of pixels with the correct label in the region.

5. The method according to claim 4, wherein, The step of dividing the sample image into multiple regions includes: The sample image is divided into multiple regions based on at least one of the semantic features, texture features, color features, and brightness features of multiple pixels in the sample image; and / or, The sample image is divided into multiple regions according to a preset size.

6. The method according to claim 1, wherein, The acquisition of the first training history corresponding to any pixel in the noise sample set includes: The deep model is trained using the noise sample set to obtain the first training history corresponding to any pixel in the noise sample set.

7. The method according to claim 6, further comprising: The original sample set is subjected to noise processing to obtain the noise sample set; Wherein, the number of pixels with incorrect labels in the original sample set is less than or equal to a first threshold, and the number of pixels with incorrect labels in the noise sample set is greater than or equal to a second threshold; The second threshold is greater than or equal to the first threshold.

8. A method for detecting noise tags, comprising: Obtain the second training history corresponding to any pixel in the image to be detected; Wherein, the image to be detected includes labeled pixels, the labels being used to characterize the category to which the pixels belong; the second training history contains a third prediction result of the category to which the pixels in the image to be detected belong; and, Based on the second training history and the noise label detector, the predicted attribute corresponding to the pixel is obtained, and the predicted attribute corresponding to the pixel is used to characterize the accuracy of the label of the pixel; The noise tag detector is generated by the method described in any one of claims 1-7.

9. The method according to claim 8, wherein, The step of obtaining the predicted attributes based on the second training history and the noise label detector includes: The image to be detected is divided into multiple regions to be detected, and each region to be detected includes multiple pixels; Based on the second training history corresponding to any pixel in the region to be detected, calculate the fourth prediction result of the category to which the region to be detected belongs; The fourth prediction result of the category to which the region to be detected belongs is input into the noise label detector to obtain the prediction attribute corresponding to the region to be detected. The prediction attribute corresponding to the region to be detected is used to characterize the accuracy of the label of the pixel in the region to be detected.

10. The method according to claim 9, wherein, The step of dividing the image to be detected into multiple regions to be detected includes: The image to be detected is divided according to at least one of the semantic features, texture features, color features, and brightness features of multiple pixels in the image to be detected, to obtain the multiple regions to be detected; and / or, The image to be detected is divided according to a preset size to obtain the multiple regions to be detected.

11. The method according to claim 8, wherein, The step of obtaining the predicted attributes based on the second training history and the noise label detector includes: The second training history corresponding to any pixel in the image to be detected is input into the noise label detector to obtain the predicted attribute corresponding to the pixel. The predicted attribute corresponding to the pixel is used to characterize the accuracy of the label of the pixel.

12. The method according to claim 8, wherein, The step of obtaining the second training history corresponding to any pixel in the image to be detected includes: The deep model is trained using the image to be detected to obtain a second training history corresponding to any pixel in the image to be detected.

13. An apparatus for generating a noise label detector, the noise label detector being used to detect pixels with incorrect labels in a sample image, the apparatus comprising: A first acquisition module is configured to acquire a first training history corresponding to any pixel in a noise sample set; wherein the noise sample set includes multiple sample images, and some or all of the sample images in the noise sample set include pixels with incorrect labels, the labels being used to characterize the category to which the pixel belongs; the first training history includes a first prediction result of the category to which the pixel in the sample image belongs; and, The generation module is used to generate a noise label detector using the first training history and the noise information of the noise sample set; The noise label detector is generated by: dividing the sample image into multiple regions, each region comprising multiple pixels; each region including a superpixel region, which is composed of pixels that are adjacent in position and have similar color, brightness, and texture; calculating a second prediction result of the category to which the region belongs based on a first training history corresponding to the pixels in the region; inputting the second prediction result into an initial noise label detector to obtain a predicted attribute corresponding to the region; determining a second true attribute corresponding to the region based on the noise information of the noise sample set; calculating a binary cross-entropy loss function corresponding to the region using the predicted attribute and the second true attribute; summing the binary cross-entropy loss functions corresponding to each region to obtain the loss function of the initial noise label detector; the predicted attribute including the proportion of incorrectly labeled pixels in the region, or the proportion of correctly labeled pixels in the region; and adjusting the initial noise label detector according to the loss function to obtain the final noise label detector.

14. A noise tag detection device, comprising: The second acquisition module is used to acquire the second training history corresponding to any pixel in the image to be detected; Wherein, the image to be detected includes labeled pixels, the labels being used to characterize the category to which the pixels belong; the second training history contains a third prediction result of the category to which the pixels in the image to be detected belong; and, The prediction module is used to obtain the predicted attribute corresponding to the pixel based on the second training history and the noise label detector, and the predicted attribute corresponding to the pixel is used to characterize the accuracy of the label of the pixel; The noise tag detector is generated by the apparatus of claim 13.

15. An electronic device comprising: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-7 or 8-12.

16. A non-transitory computer-readable storage medium storing computer instructions, wherein, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-7 or 8-12.

17. A computer program product comprising a computer program that, when executed by a processor, implements the method according to any one of claims 1-7 or 8-12.

Images