Target detection data mark noise identification method and related equipment

Abstract

The invention provides a target detection data mark noise identification method and related equipment, and relates to the technical field of computer vision. According to the invention, multi-scale target prediction is carried out on a plurality of target detection sample images with annotation boxes, so that low-dimensional feature vectors of all candidate prediction boxes corresponding to a plurality of feature scales of the plurality of target detection sample images are obtained; performing feature vector clustering on the low-dimensional feature vectors of all target prediction frames belonging to the same target category under the same feature scale of a plurality of target detection sample images to obtain feature distances between the target prediction frames and an actual clustering center; distance distribution statistics is carried out on respective feature distances of the target prediction frames, so that outlier prediction frames in the target prediction frames can be used as marked noise frames under the feature scale, and numerous marked noise in the corresponding sample data set can be quickly and accurately identified; and the data set cleaning efficiency and the model generalization capability are improved.

Description

Technical Field

[0001] This application relates to the field of computer vision technology, and more specifically, to a method and related equipment for identifying noise in target detection data labeling. Background Technology

[0002] With the continuous development of science and technology, target detection technology is being used more and more widely in various industries. Service robots (such as logistics robots and sweeping robots) are an important application of target detection technology. Target detection technology is usually used to realize the obstacle recognition function of service robots in order to improve the autonomous navigation and obstacle avoidance performance of service robots.

[0003] Currently, mainstream object detection algorithms achieve object detection functionality through lightweight model design, but their model performance is highly dependent on the annotation quality of the training sample data. However, during the data annotation process of training sample data, subjective errors in manual annotation and interference from complex environmental factors in the sample images (such as transparent objects, low lighting conditions, or dynamic obstacles) often lead to a large amount of noise in the corresponding training sample dataset (e.g., false negatives, false positives, and box shifts), which seriously affects the annotation quality of the training sample dataset. Summary of the Invention

[0004] In view of this, the purpose of this application is to provide a method, computer device and readable storage medium for identifying label noise in object detection data. This method can automatically extract all labeled sample images and perform feature clustering statistical analysis on the effective predicted boxes for a specified object category at multiple feature scales. This fully explores the complementarity of multi-scale features (including shallow detail features and deep semantic features) in the label noise detection process, captures abnormal predicted boxes in the deep feature space, and thus quickly and accurately identifies the numerous label noises actually contained in the corresponding sample dataset. This facilitates the improvement of sample dataset cleaning efficiency and the generalization ability of the object detection model (which is trained based on the cleaned sample dataset).

[0005] To achieve the above objectives, the technical solutions adopted in the embodiments of this application are as follows: In a first aspect, this application provides a method for identifying noise in target detection data, the method comprising: Obtain a target detection dataset for at least one target category, wherein the target detection dataset includes multiple target detection sample images, each with all labeled bounding boxes corresponding to a target category; Multi-scale target prediction is performed on the target detection dataset to obtain low-dimensional feature vectors of all candidate prediction boxes corresponding to each of the multiple target detection sample images at multiple feature scales, wherein all candidate prediction boxes of the same target detection sample image at a single feature scale correspond to a target category. For the multiple target detection sample images, feature vector clustering is performed on the low-dimensional feature vectors of all target prediction boxes belonging to the same target category at the same feature scale to obtain the feature distance between each target prediction box and the actual cluster center. Distance distribution statistics are performed on the feature distances of all the target prediction boxes to obtain the outlier prediction boxes among all the target prediction boxes, and the outlier prediction boxes are used as the labeled noise boxes of the corresponding target detection sample images.

[0006] In an optional implementation, the step of performing multi-scale target prediction on the target detection dataset to obtain low-dimensional feature vectors of all candidate prediction boxes corresponding to each of the multiple target detection sample images at multiple feature scales includes: For each object detection sample image, a pre-trained multi-scale object prediction model is called to extract the feature maps corresponding to the object detection sample image at multiple feature scales, and to predict the initial prediction box, prediction confidence and prediction category of each pixel in each feature map of the object detection sample image. Based on the positions of all bounding boxes in the target detection sample image and the prediction confidence of each pixel in a single feature map, invalid bounding boxes are filtered out from all the initial predicted bounding boxes in the feature map of the target detection sample image to obtain all the candidate predicted bounding boxes in the feature map of the target detection sample image. Extract the low-dimensional feature vectors of all candidate predicted boxes in the feature map of the target detection sample image.

[0007] In an optional implementation, the step of filtering out invalid bounding boxes from all initial predicted bounding boxes in the feature map of the target detection sample image based on the positions of all labeled bounding boxes in the target detection sample image and the prediction confidence of each pixel in a single feature map, to obtain all candidate predicted bounding boxes in the feature map of the target detection sample image, includes: For each bounding box in the target detection sample image, determine all candidate pixels whose corresponding pixel mapping position in a single feature map of the target detection sample image is within the bounding box; Among all the candidate pixels, select all target pixels whose corresponding prediction confidence exceeds a preset confidence threshold; Non-maximum suppression is applied to the initial prediction boxes of all target pixels to obtain the candidate prediction boxes corresponding to the labeled boxes in the feature map of the target detection sample image.

[0008] In an optional implementation, the step of extracting the low-dimensional feature vectors of each candidate predicted bounding box in the feature map from the target detection sample image includes: For each candidate prediction box in the feature map of the target detection sample image, extract the local region features of the candidate prediction box in the feature map. The local region features corresponding to the candidate prediction box are vectorized into low dimensions to obtain the low-dimensional feature vector of the candidate prediction box at the corresponding feature scale.

[0009] In an optional implementation, the step of performing low-dimensional vectorization processing on the local region features corresponding to the candidate prediction box to obtain the low-dimensional feature vector of the candidate prediction box at the corresponding feature scale includes: Global average pooling is performed on the local region features corresponding to the candidate prediction box to obtain the global features of all channels of the local region features; Channel attention weights are learned based on the global features of all channels to obtain a channel attention weight matrix that matches the local region features; The channel attention weight matrix and the local region features are multiplied channel by channel to obtain the low-dimensional feature vector of the candidate prediction box.

[0010] In an optional implementation, the step of performing distance distribution statistics on the feature distances of each of the target prediction boxes to obtain outlier prediction boxes among all the target prediction boxes includes: Based on the feature distances of all target prediction boxes belonging to the same target category at the same feature scale, probability density histograms are plotted to obtain the corresponding target probability density curves. In the target probability density curve, determine the center value of the target distance interval corresponding to the main peak of the curve, and determine all outlier distance intervals of the target probability density curve based on the center value of the target distance interval and a preset outlier ratio threshold; The target prediction box whose corresponding feature distance is within any outlier distance interval is taken as the outlier prediction box.

[0011] In an optional implementation, the step of determining all outlier distance intervals of the target probability density curve based on the center value of the target distance interval and a preset outlier proportion threshold includes: For each statistically analyzed distance interval involved in the target probability density curve, calculate the absolute value of the distance difference between the center value of the statistically analyzed distance interval and the center value of the target distance interval, and calculate the actual ratio between the absolute value of the distance difference and the center value of the target distance interval. Detect whether the actual ratio value is greater than or equal to the preset outlier ratio threshold; If the actual proportion value is detected to be greater than or equal to the preset outlier proportion threshold, the statistically calculated distance interval is taken as an outlier distance interval.

[0012] In an optional implementation, the method further includes: The target detection sample images with marked noise boxes are sent to the data annotation review terminal for manual quality inspection.

[0013] Secondly, this application provides a computer device including a processor and a memory, wherein the memory stores a computer program that can be executed by the processor, and the processor can execute the computer program to implement the target detection data tag noise recognition method described in any of the foregoing embodiments.

[0014] Thirdly, this application provides a readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a computer device, it implements the target detection data labeling noise recognition method described in any of the foregoing embodiments.

[0015] In this case, the beneficial effects of the embodiments of this application may include the following: This application performs multi-scale target prediction on multiple target detection sample images labeled with bounding boxes in an object detection dataset. This yields low-dimensional feature vectors of all candidate predicted bounding boxes for each of the multiple target detection sample images at various feature scales. Then, feature vector clustering is performed on the low-dimensional feature vectors of all target predicted bounding boxes belonging to the same target category at the same feature scale. This yields the feature distances between each of these target predicted bounding boxes and the actual cluster centers. Distance distribution statistics are then performed on these feature distances to identify outlier predicted bounding boxes as labeled noise boxes in the corresponding target detection sample images at that feature scale. This approach automatically extracts all labeled sample images (i.e., multiple object detection sample images in the object detection dataset) and performs feature clustering statistical analysis on the effective predicted boxes (i.e., object predicted boxes) for a specified object category at multiple feature scales. This fully leverages the complementarity of the multi-scale features (including shallow detail features and deep semantic features) of each labeled sample image in the label noise detection process. It facilitates the capture of anomalous predicted boxes (i.e., outlier predicted boxes) in the deep feature space of the corresponding labeled sample images, thereby quickly and accurately identifying the numerous label noises actually contained in the corresponding sample dataset. This improves the efficiency of sample dataset cleaning and the generalization ability of the object detection model.

[0016] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A schematic diagram of the composition of a computer device provided in the embodiments of this application; Figure 2 This is one of the flowcharts illustrating the target detection data labeling and noise recognition method provided in the embodiments of this application; Figure 3 for Figure 2 A flowchart illustrating the sub-steps included in step S220; Figure 4 for Figure 2 A flowchart illustrating the sub-steps included in step S240; Figure 5This is a schematic diagram of the probability density histogram provided for multiple target prediction boxes with the same feature scale and target category in an embodiment of this application; Figure 6 This is the second flowchart illustrating the target detection data labeling noise recognition method provided in the embodiments of this application.

[0019] Icons: 10-Computer equipment; 11-Memory; 12-Processor; 13-Communication unit. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0021] Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0022] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0023] In the description of this application, it should be understood that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product is in use, or the orientation or positional relationship commonly understood by those skilled in the art. They are used only for the convenience of describing this application and simplifying the description, and are not intended to indicate or imply that the equipment or component referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application.

[0024] In the description of this application, it should also be noted that, unless otherwise expressly specified and limited, the terms "set up," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0025] Furthermore, it is understood in the description of this application that relational terms such as "first" and "second" are used merely 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. Moreover, 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. Those skilled in the art will understand the specific meaning of the above terms in this application based on the specific circumstances.

[0026] Through painstaking research, the applicant discovered that current methods for identifying labeled noise in target detection sample datasets mainly employ the following approaches: (1) rule-based filtering methods relying on manual quality checks; and (2) automated noise screening methods relying on low-dimensional statistics (e.g., bounding box confidence or bounding box intersection-union ratio). However, it is worth noting that these two approaches are inherently difficult to adapt to complex and varied noise types. They rely solely on statistical surface image features for noise identification, neglecting the semantic information of deep image features, resulting in a high false positive rate in the final noise identification results. This severely restricts the effectiveness of training sample data cleaning and the improvement of the generalization ability of target detection models.

[0027] In this context, to address the aforementioned issues, this application provides a method for identifying label noise in target detection data, a computer device, and a readable storage medium. This method fully leverages the complementarity of multi-scale features (including shallow detail features and deep semantic features) of numerous labeled sample images (i.e., multiple target detection sample images labeled with bounding boxes, where each bounding box corresponds to any specified target type) during the label noise detection process. This facilitates the capture of abnormal prediction boxes in the deep feature space of each labeled sample image, thereby quickly and accurately identifying the numerous label noises actually contained in the corresponding sample dataset, improving the efficiency of sample dataset cleaning and the generalization ability of the target detection model.

[0028] The following detailed description of some embodiments of this application is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0029] Please refer to Figure 1 , Figure 1 This is a schematic diagram of the computer device 10 provided in this application embodiment. In this application embodiment, the computer device 10 can communicate with a data annotation review terminal held by a data annotation expert, and obtain the target detection dataset provided by the data annotation expert through the data annotation review terminal. Then, it performs fast and accurate label noise recognition processing on the target detection dataset, which facilitates the improvement of the dataset cleaning efficiency and thus improves the model generalization ability of the corresponding target detection model. The target detection dataset is the sample dataset used to train the target detection model. The target detection dataset consists of multiple target detection sample images. Each target detection sample image is manually labeled with at least one annotation box by a data annotation expert. Each annotation box corresponds to a single target category (i.e., target object category). A single target category can correspond to multiple annotation boxes in the same target detection sample image. The computer device 10 and the data annotation review terminal can be, but are not limited to, personal computers, cloud servers, laptops, tablets, etc. In one embodiment of this example, the target detection dataset may be a sample dataset configured for the obstacle recognition function of the robotic vacuum cleaner. In this case, all target categories involved in the target detection dataset belong to the obstacle types that the robotic vacuum cleaner can photograph within its effective field of view.

[0030] In this embodiment of the application, the computer device 10 may include a memory 11, a processor 12, and a communication unit 13. The memory 11, the processor 12, and the communication unit 13 are electrically connected to each other directly or indirectly to achieve data transmission or interaction. For example, these components can be electrically connected to each other via one or more communication buses or signal lines.

[0031] In this embodiment, the memory 11 may be, but is not limited to, random access memory (RAM), read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), etc. The memory 11 is used to store computer programs, and the processor 12 can execute the computer programs accordingly after receiving execution instructions.

[0032] Furthermore, the memory 11 is also used to store a pre-trained multi-scale target prediction model. The multi-scale target prediction model is used to perform bounding box prediction processing on the same labeled sample image (i.e., target detection sample image) at multiple feature scales (including surface feature scale and deep feature scale) to obtain the target prediction information of each bounding box in the image feature map of the corresponding labeled sample image at different feature scales (wherein the target prediction information of a single bounding box includes the bounding box coordinates in the image feature map of the corresponding feature scale, the prediction category (i.e., the specific target category) of the bounding box, and the prediction confidence of the bounding box belonging to the corresponding prediction category). The multi-scale target prediction model can be implemented using target detection model architectures such as the YOLO series models (e.g., YOLOv6 model, YOLOv8 model) and the SSD series models with anchorless bounding box detection mechanisms.

[0033] In this embodiment, the processor 12 can be an integrated circuit chip with signal processing capabilities. The processor 12 can be a general-purpose processor, including at least one of a Central Processing Unit (CPU), Graphics Processing Unit (GPU), Network Processor (NP), Digital Signal Processor (DSP), Application-Specific Integrated Circuit (ASIC), Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor can be a microprocessor or any conventional processor, capable of implementing or executing the methods, steps, and logic block diagrams disclosed in this embodiment.

[0034] In this embodiment of the application, the communication unit 13 is used to establish a communication connection between the computer device 10 and other electronic devices through a network, and to send and receive data through the network, wherein the network includes a wired communication network and a wireless communication network.

[0035] In this embodiment, the computer device 10 may pre-store a specific computer program related to the target detection data label noise recognition function in the memory 11, and automatically extract all labeled sample images in the target detection sample dataset (i.e., the target detection dataset mentioned above) and perform feature clustering statistical analysis on the effective prediction boxes for the specified target category at multiple feature scales, so as to fully explore the complementarity of multi-scale features in the label noise detection process, capture abnormal prediction boxes in the deep feature space, and thus quickly and accurately identify the numerous label noises actually contained in the corresponding sample dataset, which is conducive to improving the sample dataset cleaning efficiency and the model generalization ability of the target detection model.

[0036] Understandable Figure 1 The block diagram shown is only a schematic diagram of one configuration of the computer device 10. The computer device 10 may also include components such as... Figure 1 The more or fewer components shown, or having the same Figure 1 The different configurations shown. Figure 1 The components shown can be implemented using hardware, software, or a combination thereof.

[0037] In this application, to ensure that the computer device 10 can perform fast and accurate label noise recognition processing on the target detection dataset at multiple feature scales, thereby improving the dataset cleaning efficiency and the generalization ability of the corresponding target detection model, this application provides a method for label noise recognition in target detection data to achieve the aforementioned objectives. The method for label noise recognition in target detection data provided in this application will be described in detail below.

[0038] Please refer to Figure 2 , Figure 2 This is one of the flowcharts illustrating the target detection data labeling and noise recognition method provided in this application embodiment. In this application embodiment, Figure 2 The target detection data labeling noise identification method shown may include steps S210 to S240.

[0039] Step S210: Obtain a target detection dataset for at least one target category, wherein the target detection dataset includes multiple target detection sample images, each with its own bounding boxes corresponding to a target category.

[0040] In this embodiment, a single target detection sample image may be marked with multiple bounding boxes corresponding to the same target category; for each target detection sample image, the target detection dataset records the bounding box coordinates of each bounding box in the target detection sample image, as well as the actual target category to which each bounding box belongs.

[0041] Step S220: Perform multi-scale target prediction on the target detection dataset to obtain low-dimensional feature vectors of all candidate prediction boxes corresponding to each of the multiple target detection sample images at multiple feature scales.

[0042] In this embodiment, after obtaining the target detection dataset, the computer device 10 extracts feature maps for each target detection sample image in the dataset at multiple feature scales. It then performs bounding box prediction processing on the feature maps corresponding to each of the extracted feature scales to obtain target prediction information (including the bounding box coordinates of the initial prediction box in the corresponding feature map, the predicted category (i.e., the specific target category) of the initial prediction box, and the prediction confidence of the prediction box belonging to the corresponding prediction category) for each pixel in different feature maps (i.e., feature maps at different feature scales). Then, through invalid box filtering, it further determines candidate prediction boxes with a high probability of reliability in different feature maps (wherein, the set of candidate prediction boxes for the same target detection sample image at different feature maps can be partially the same or completely identical). Simultaneously, it extracts the low-dimensional feature vectors of all candidate prediction boxes within the same feature map. Specifically, all candidate prediction boxes of the same target detection sample image at a single feature scale correspond to one target category; the set of candidate prediction boxes of the same target detection sample image at different feature scales can be partially the same or completely identical.

[0043] Alternatively, please refer to Figure 3 , Figure 3 yes Figure 2 A flowchart illustrating the sub-steps included in step S220. In this embodiment of the application, step S220 may include sub-steps S221 to S223 to automatically extract low-dimensional representations (i.e., low-dimensional feature vectors) of target object features from all labeled sample images at multiple feature scales for the specified target category, representing effective prediction boxes (i.e., candidate prediction boxes) of the target object.

[0044] Sub-step S221: For each object detection sample image, call the pre-trained multi-scale object prediction model to extract the feature maps corresponding to the object detection sample image at multiple feature scales, and predict the initial prediction box, prediction confidence and prediction category of each pixel in each feature map of the object detection sample image.

[0045] Sub-step S222: Based on the positions of all the labeled bounding boxes in the target detection sample image and the prediction confidence of each pixel in a single feature map, invalid bounding boxes are filtered out from all the initial predicted bounding boxes in the feature map of the target detection sample image to obtain all the candidate predicted bounding boxes in the feature map of the target detection sample image.

[0046] In this embodiment, after the computer device 10 determines the target prediction information of all initial prediction boxes (i.e., the initial prediction boxes of each pixel) in the feature map of a target detection sample image at a certain feature scale, it will filter out the initial prediction boxes that are unrelated to the labeled boxes of the target detection sample image and filter out the initial prediction boxes that are related to the labeled boxes of the target detection sample image but have low reliability, based on the positions of all labeled boxes in the target detection sample image (i.e., the bounding box coordinates of each labeled box) and the prediction confidence of each initial prediction box in the feature map, thereby selecting all candidate prediction boxes of the target detection sample image in the feature map.

[0047] Optionally, in one embodiment of this example, the above sub-step S222 may include: For each bounding box in the target detection sample image, determine all candidate pixels whose corresponding pixel mapping position in a single feature map of the target detection sample image is within the bounding box; Among all the candidate pixels, select all target pixels whose corresponding prediction confidence exceeds a preset confidence threshold; Non-maximum suppression is applied to the initial prediction boxes of all target pixels to obtain the candidate prediction boxes corresponding to the labeled boxes in the feature map of the target detection sample image.

[0048] Sub-step S224: Extract the low-dimensional feature vectors of all candidate prediction boxes in the feature map of the target detection sample image.

[0049] In this embodiment, after the computer device 10 determines all candidate prediction boxes in the feature map of a target detection sample image at a certain feature scale, it can extract the local region features of each candidate prediction box in the feature map and perform low-dimensional vectorization processing on the local region features corresponding to the candidate prediction box to obtain a low-dimensional representation (i.e., low-dimensional feature vector) of the target object features at the corresponding feature scale. The computer device 10 can implement the low-dimensional feature vectorization function based on any one of the feature dimensionality reduction mechanisms, such as channel attention, low-dimensional embedding, and principal component analysis (PCA).

[0050] Optionally, in one embodiment of this invention, the computer device 10 can implement the feature low-dimensional vectorization function based on the channel attention mechanism. In this case, the above step "perform low-dimensional vectorization processing on the local region features corresponding to the candidate prediction box to obtain the low-dimensional feature vector of the candidate prediction box at the corresponding feature scale" may include: Global average pooling is performed on the local region features corresponding to the candidate prediction box to obtain the global features of all channels of the local region features (including the global features corresponding to the local region features at all feature channels involved). Channel attention weights are learned based on the global features of all channels to obtain a channel attention weight matrix that matches the local region features; The channel attention weight matrix and the local region features are multiplied channel by channel to obtain the low-dimensional feature vector of the candidate prediction box.

[0051] Therefore, by executing the above sub-steps S221 to S223, this application can automatically extract low-dimensional representations of target object features of all labeled sample images at multiple feature scales for the specified target category, namely, the effective prediction boxes (i.e., candidate prediction boxes).

[0052] Step S230: Perform feature vector clustering on the low-dimensional feature vectors of all target prediction boxes belonging to the same target category under the same feature scale in multiple target detection sample images to obtain the feature distance between each target prediction box and the actual cluster center.

[0053] In this embodiment, after the computer device 10 determines the low-dimensional feature vectors of all candidate predicted boxes at different feature scales for each of the multiple target detection sample images, it divides the target predicted box sets according to the target category and feature scale, so that all target predicted boxes in each set belong to the same target category at the same feature scale. Then, for each set of target predicted boxes, the computer device 10 performs feature vector clustering on the low-dimensional feature vectors of all target predicted boxes in the set to determine the low-dimensional feature vector of the actual cluster center of the set. Then, by calculating the Euclidean distance between the low-dimensional feature vectors of all target predicted boxes in the set and the low-dimensional feature vector of the corresponding actual cluster center, the feature distance between each target predicted box in the set and the actual cluster center is obtained.

[0054] Step S240: Perform distance distribution statistics on the feature distances of all target prediction boxes to obtain outlier prediction boxes among all target prediction boxes, and use the outlier prediction boxes as labeled noise boxes of the corresponding target detection sample images.

[0055] In this embodiment, after the computer device 10 determines the feature distances of all target prediction boxes in a certain target prediction box set, it performs mathematical statistical analysis (which may include, but is not limited to, variance analysis, probability density distribution analysis, histogram distribution analysis, etc.) on the feature distances of all target prediction boxes in the target prediction box set to identify outlier prediction boxes in the target prediction box set whose feature distances differ too much from those of other target prediction boxes. At this time, the outlier prediction box can serve as a labeled noise box representing the labeled noise of the target detection sample image at the corresponding feature scale (i.e., the feature scale associated with the target prediction box set). This enables feature clustering statistical analysis of all labeled sample images for the same target category at the same feature scale, so as to fully explore the complementarity of multi-scale features (including shallow detail features and deep semantic features) in the label noise detection process. It can effectively capture abnormal prediction boxes in the deep feature space, and thus quickly and accurately identify all the labeled noise actually contained in the target detection data at different feature scale levels. This facilitates the improvement of sample dataset cleaning efficiency and the generalization ability of the target detection model.

[0056] Optionally, please refer to the following: Figure 4 and Figure 5 ,in Figure 4 yes Figure 2 A flowchart illustrating the sub-steps included in step S240. Figure 5 This is a schematic diagram of probability density histogram drawing provided by an embodiment of this application for multiple target prediction boxes with the same feature scale and target category. In this embodiment, the computer device 10 can use mathematical statistical analysis methods combining probability density distribution analysis and histogram distribution analysis to achieve accurate identification of all outlier prediction boxes within a single target prediction box set. At this time, the step "perform distance distribution statistics on the feature distances of all target prediction boxes to obtain outlier prediction boxes among all target prediction boxes" in step S240 may include sub-steps S241 to S243.

[0057] Sub-step S241: Draw a probability density histogram based on the feature distances of all target prediction boxes belonging to the same target category under the same feature scale, and obtain the corresponding target probability density curve.

[0058] In this embodiment, when the computer device 10 draws a probability density histogram for a single set of target prediction boxes, it constructs multiple feature distance intervals in a continuous distribution starting from 0 according to a specific feature distance unit interval value (e.g., 0.3). The distance difference between the upper and lower limits of each feature distance interval is consistent with the feature distance unit interval value. Then, it counts the frequency of occurrence of the feature distances of all target prediction boxes within the set within these multiple feature distance intervals. Next, for each counted distance interval (i.e., the feature distance interval containing the feature distance of at least one target prediction box), it calculates the probability density value of that counted distance interval based on the frequency statistics. Finally, it performs histogram drawing on the probability density values ​​of all counted distance intervals to obtain the following result: Figure 5 The distance histogram shown (its distance in) Figure 5 (Using blue rectangular blocks as representations), the distance histogram is then fitted with a probability density function to obtain, as shown below. Figure 5 The target probability density curve shown (which is in) Figure 5 (The red curve is used to represent this).

[0059] Sub-step S242: Determine the center value of the target distance interval corresponding to the main peak value of the target probability density curve, and determine all outlier distance intervals of the target probability density curve based on the center value of the target distance interval and the preset outlier ratio threshold. In this embodiment, the main peak of the curve is the maximum probability density value of the corresponding target probability density curve (which is composed of...). Figure 5 The curve's main peak calibration line (represented by a green dashed line) intersects with the target probability density curve to obtain the target distance interval center value, which is the abscissa value of the target probability density curve corresponding to the main peak value of the curve (i.e., Figure 5 The intersection of the main peak calibration line of the curve and the horizontal coordinate axis); after obtaining the center value of the target distance interval of a single target prediction box set, the computer device 10 will select the corresponding distance length exceeding the distance difference magnitude threshold (which is obtained by multiplying the center value of the target distance interval and the preset outlier ratio threshold) based on the center values ​​of all statistically calculated distance intervals of the target prediction box set at the corresponding target probability density curve. Figure 5 In the model, each statistically calculated distance interval can be represented by an outlier partition boundary line (represented by a purple dashed line) as the outlier distance interval (i.e., Figure 5 The feature distance intervals corresponding to the multiple blue rectangular blocks located to the right of the right outlier boundary line (i.e., the outlier boundary line far from the origin of the coordinate system).

[0060] Optionally, in one embodiment of this example, the step "determining all outlier distance intervals of the target probability density curve based on the target distance interval center value and the preset outlier ratio threshold" in sub-step S242 may include: For each statistically analyzed distance interval involved in the target probability density curve, calculate the absolute value of the distance difference between the center value of the statistically analyzed distance interval and the center value of the target distance interval, and calculate the actual ratio between the absolute value of the distance difference and the center value of the target distance interval. Detect whether the actual ratio value is greater than or equal to the preset outlier ratio threshold; If the actual proportion value is detected to be greater than or equal to the preset outlier proportion threshold, the statistically calculated distance interval is taken as an outlier distance interval.

[0061] Sub-step S243: Take the target prediction box whose corresponding feature distance is within any outlier distance interval as the outlier prediction box.

[0062] Therefore, by executing the above sub-steps S241 to S243, this application can achieve the function of accurately identifying all outlier prediction boxes within a single target prediction box set by using mathematical statistical analysis methods that combine probability density distribution analysis and histogram distribution analysis.

[0063] Meanwhile, by executing the above steps S210 to S240, this application can perform feature clustering statistical analysis on the effective prediction boxes of the same target category for all labeled sample images at different feature scales, so as to fully explore the complementarity of multi-scale features in the label noise detection process and effectively capture abnormal prediction boxes in the deep feature space. This allows for the rapid and accurate identification of all the label noise actually contained in the target detection data at different feature scales, which is conducive to improving the sample dataset cleaning efficiency and the generalization ability of the target detection model.

[0064] Alternatively, please refer to Figure 6 , Figure 6 This is a second schematic flowchart of the target detection data labeling noise recognition method provided in this application embodiment. In this application embodiment, with Figure 2 Compared to the target detection data labeling and noise recognition method shown, Figure 6 The target detection data labeling noise identification method shown may also include step S250, so that data labeling experts can promptly correct the labeled sample images involving labeling noise, thereby improving the labeling quality and cleaning efficiency of the target detection dataset.

[0065] Step S250: Send the target detection sample image with the marked noise box to the data annotation review terminal for manual quality inspection.

[0066] In this embodiment, when the computer device 10 determines that a target detection sample image contains labeled noise boxes (i.e., outlier prediction boxes) at a certain feature scale level, the computer device 10 uses a different border color than the labeled border color of the target detection sample image (where multiple feature scales may correspond to the same border color, or multiple feature scales may correspond to different border colors). The computer device 10 then performs border annotation processing in the target detection sample image according to the target prediction information corresponding to the outlier prediction box, thereby automatically annotating the labeled noise boxes involved in the target detection sample image. After the computer device 10 has annotated all the labeled noise boxes of a target detection sample image at different feature scale levels, the computer device 10 sends the target detection sample image to the data annotation review terminal. This allows the data annotation experts at the data annotation review terminal to promptly correct the existing labeled boxes of the target detection sample image. The data annotation experts then update the corrected target detection sample image into the target detection dataset, thereby improving the annotation quality and cleaning efficiency of the target detection dataset.

[0067] Therefore, by performing the above step S250, this application enables data annotation experts to promptly correct the annotations of labeled sample images involving label noise, thereby improving the annotation quality and cleaning efficiency of the target detection dataset.

[0068] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0069] Furthermore, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part. If the various functions provided in this application are implemented in the form of software functional modules and sold or used as independent products, they can be stored in a storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a readable storage medium and includes several instructions to cause a computer device (e.g., a laptop, a robot vacuum cleaner, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned readable storage medium includes: USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, optical disks, and other media capable of storing program code.

[0070] The above descriptions are merely various embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for identifying noise in target detection data, characterized in that, The method includes: Obtain a target detection dataset for at least one target category, wherein the target detection dataset includes multiple target detection sample images, each with all labeled bounding boxes corresponding to a target category; Multi-scale target prediction is performed on the target detection dataset to obtain low-dimensional feature vectors of all candidate prediction boxes corresponding to each of the multiple target detection sample images at multiple feature scales, wherein all candidate prediction boxes of the same target detection sample image at a single feature scale correspond to a target category. For the multiple target detection sample images, feature vector clustering is performed on the low-dimensional feature vectors of all target prediction boxes belonging to the same target category at the same feature scale to obtain the feature distance between each target prediction box and the actual cluster center. Distance distribution statistics are performed on the feature distances of all the target prediction boxes to obtain the outlier prediction boxes among all the target prediction boxes, and the outlier prediction boxes are used as the labeled noise boxes of the corresponding target detection sample images.

2. The method according to claim 1, characterized in that, The step of performing multi-scale target prediction on the target detection dataset to obtain low-dimensional feature vectors of all candidate prediction boxes corresponding to each of the multiple target detection sample images at multiple feature scales includes: For each object detection sample image, a pre-trained multi-scale object prediction model is called to extract the feature maps corresponding to the object detection sample image at multiple feature scales, and the initial prediction box, prediction confidence and prediction category of each pixel in each feature map of the object detection sample image are predicted. Based on the positions of all the bounding boxes in the target detection sample image and the prediction confidence of each pixel in a single feature map, invalid bounding boxes are filtered out from all the initial predicted bounding boxes in the feature map of the target detection sample image to obtain all the candidate predicted bounding boxes in the feature map of the target detection sample image. Extract the low-dimensional feature vectors of all candidate predicted boxes in the feature map of the target detection sample image.

3. The method according to claim 2, characterized in that, The step of filtering out invalid bounding boxes from all initial predicted bounding boxes in the feature map of the target detection sample image based on the positions of all labeled bounding boxes in the target detection sample image and the prediction confidence of each pixel in a single feature map, to obtain all candidate predicted bounding boxes in the feature map of the target detection sample image, includes: For each bounding box in the target detection sample image, determine all candidate pixels whose corresponding pixel mapping position in a single feature map of the target detection sample image is within the bounding box; Among all the candidate pixels, select all target pixels whose corresponding prediction confidence exceeds a preset confidence threshold; Non-maximum suppression is applied to the initial prediction boxes of all target pixels to obtain the candidate prediction boxes corresponding to the labeled boxes in the feature map of the target detection sample image.

4. The method according to claim 2, characterized in that, The step of extracting the low-dimensional feature vectors of all candidate prediction boxes in the feature map of the target detection sample image includes: For each candidate prediction box in the feature map of the target detection sample image, extract the local region features of the candidate prediction box in the feature map. The local region features corresponding to the candidate prediction box are vectorized into low dimensions to obtain the low-dimensional feature vector of the candidate prediction box at the corresponding feature scale.

5. The method according to claim 4, characterized in that, The step of performing low-dimensional vectorization processing on the local region features corresponding to the candidate prediction box to obtain the low-dimensional feature vector of the candidate prediction box at the corresponding feature scale includes: Global average pooling is performed on the local region features corresponding to the candidate prediction box to obtain the global features of all channels of the local region features; Channel attention weights are learned based on the global features of all channels to obtain a channel attention weight matrix that matches the local region features; The channel attention weight matrix and the local region features are multiplied channel by channel to obtain the low-dimensional feature vector of the candidate prediction box.

6. The method according to any one of claims 1-5, characterized in that, The step of performing distance distribution statistics on the feature distances of each of the target prediction boxes to obtain outlier prediction boxes among all the target prediction boxes includes: Based on the feature distances of all target prediction boxes belonging to the same target category at the same feature scale, probability density histograms are plotted to obtain the corresponding target probability density curves. In the target probability density curve, determine the center value of the target distance interval corresponding to the main peak of the curve, and determine all outlier distance intervals of the target probability density curve based on the center value of the target distance interval and a preset outlier ratio threshold; The target prediction box whose corresponding feature distance is within any outlier distance interval is taken as the outlier prediction box.

7. The method according to claim 6, characterized in that, The step of determining all outlier distance intervals of the target probability density curve based on the center value of the target distance interval and a preset outlier ratio threshold includes: For each statistically analyzed distance interval involved in the target probability density curve, calculate the absolute value of the distance difference between the center value of the statistically analyzed distance interval and the center value of the target distance interval, and calculate the actual ratio between the absolute value of the distance difference and the center value of the target distance interval. Detect whether the actual ratio value is greater than or equal to the preset outlier ratio threshold; If the actual proportion value is detected to be greater than or equal to the preset outlier proportion threshold, the statistically calculated distance interval is taken as an outlier distance interval.

8. The method according to claim 1, characterized in that, The method further includes: The target detection sample images with marked noise boxes are sent to the data annotation review terminal for manual quality inspection.

9. A computer device, characterized in that, The device includes a processor and a memory, the memory storing a computer program executable by the processor, the processor executing the computer program to implement the target detection data tagging noise recognition method according to any one of claims 1-8.

10. A readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a computer device, it implements the target detection data labeling noise recognition method according to any one of claims 1-8.

Images