Methods, apparatuses, and computer program products for filtering a filtered image
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NTT DOCOMO INC
- Filing Date
- 2024-12-03
- Publication Date
- 2026-06-05
Smart Images
Figure CN122156842A_ABST
Abstract
Description
Technical Field
[0001] This disclosure relates to data processing, and more particularly to a method, apparatus, and computer program product for filtering images to be filtered. Background Technology
[0002] Object detection is a crucial technique in computer vision, designed to identify and locate one or more target objects within an image or video frame. Its applications are wide-ranging, including autonomous driving, face recognition, and the detection of diseased leaves. In object detection tasks, bounding boxes are key tools for determining the location of target objects in an image; they are rectangular regions used to identify the target's position. Bounding boxes can be determined manually by marking them on the image, or automatically using methods such as edge detection, clustering algorithms, or rule-based approaches.
[0003] With the rapid development of deep learning, object detection algorithms based on deep learning have made significant progress. YOLO (You Only Look Once) and Faster R-CNN (Region-based Convolutional Neural Network) are two popular deep learning object detection models. Deep learning models rely on large amounts of labeled data for training, and often suffer from poor performance due to insufficient labeled data or imbalanced data distribution.
[0004] To achieve better object detection results, the generalization ability of the model can be improved by expanding / augmenting the training dataset. Generative models are a type of machine learning model. They can learn the distribution of existing data and generate new data from it, making them very useful in tasks such as image generation, text generation, and audio generation. Diffusion models and GANs (Generative Adversarial Networks) are common examples of generative models. Summary of the Invention
[0005] This disclosure relates to a method, apparatus, and computer program product for filtering images to be filtered. It determines the similarity of one or more bounding box images corresponding to one or more bounding box images included in the image to be filtered, based on a reference set of bounding box images, and then determines whether to filter the image to be filtered based on these similarities. Since the reference set of bounding box images includes more bounding box images than the true set of bounding box images, more images beneficial to downstream tasks (such as various machine learning or neural network models of the aforementioned deep learning object detection model) can be selected for training in the downstream task instead of being filtered out. This increases the diversity of the filtered images, reduces overfitting, and thus improves the performance of the downstream task.
[0006] According to one aspect of this disclosure, a method for filtering an image to be filtered is provided. The method includes: deriving a set of reference bounding box images based on a set of ground truth bounding box images, wherein the reference bounding box images in the set of reference bounding box images include ground truth bounding box images from the set of ground truth bounding box images and augmented bounding box images different from the ground truth bounding box images; calculating the similarity between each bounding box image in one or more bounding box images included in the image to be filtered and the set of reference bounding box images to obtain one or more bounding box image similarities corresponding to the one or more bounding box images; and determining whether to filter the image to be filtered based on the one or more bounding box image similarities.
[0007] According to some embodiments of this disclosure, calculating the similarity between each bounding box image and a set of reference bounding box images to obtain a bounding box image similarity corresponding to the bounding box image includes: calculating the similarity between the bounding box image and each reference bounding box image in the set of reference bounding box images to obtain a plurality of first similarities; and setting the maximum similarity among the plurality of first similarities as the bounding box image similarity corresponding to the bounding box image.
[0008] According to some embodiments of this disclosure, determining whether to filter an image based on the similarity of one or more bounding box images includes: filtering the image based on the fact that the similarity of one or more bounding box images is less than a first threshold.
[0009] According to some embodiments of this disclosure, determining whether to filter an image based on one or more bounding box image similarities includes: filtering the image based on at least one of the one or more bounding box image similarities being less than a second threshold.
[0010] According to some embodiments of this disclosure, deriving a reference bounding box image set based on a set of real bounding box images includes: using the mixup blending method to derive additional bounding box images as augmented bounding box images based on real bounding box images in the set of real bounding box images.
[0011] According to some embodiments of this disclosure, based on real bounding box images in a set of real bounding box images, an additional bounding box image is derived as an augmented bounding box image using the mixup blending method, including: performing pixel-level blending or feature-level blending on the real bounding box images using the mixup blending method to derive the additional bounding box image as an augmented bounding box image.
[0012] According to some embodiments of this disclosure, based on real bounding box images in a set of real bounding box images, an additional bounding box image is derived as an augmented bounding box image using the mixup blending method, including: using the mixup blending method to blend multiple real bounding box images in the set of real bounding box images with a similarity higher than a third threshold to derive the additional bounding box image as an augmented bounding box image.
[0013] According to some embodiments of this disclosure, based on real bounding box images in a set of real bounding box images, an additional bounding box image is derived as an augmented bounding box image using the mixup blending method, including: using the mixup blending method to blend multiple real bounding box images in the set of real bounding box images with a similarity lower than a fourth threshold to derive the additional bounding box image as an augmented bounding box image.
[0014] According to some embodiments of this disclosure, the image to be filtered is an image generated by a generative model.
[0015] According to another aspect of this disclosure, an apparatus for filtering images is provided. The apparatus includes: a processor; and a memory storing one or more computer programs, wherein the one or more computer programs, when executed by the processor, cause the processor to perform the above-described method for filtering images.
[0016] According to another aspect of this disclosure, a computer program product is provided having instructions stored thereon that, when executed by a processor, cause the processor to perform the above-described method for filtering images.
[0017] According to embodiments of this disclosure, one or more bounding box images corresponding to one or more bounding box images included in the image to be filtered can be determined based on a reference bounding box image set, and then a decision on whether to filter the image to be filtered can be made based on these similarities. Since the reference bounding box image set includes more bounding box images than the true bounding box image set, more images beneficial to downstream tasks can be selected for training of downstream tasks instead of being filtered out, thereby increasing the diversity of the filtered images, reducing overfitting, and ultimately improving the performance of downstream tasks. Attached Figure Description
[0018] The aspects, features, and advantages of this disclosure will become clearer and more readily understood from the following description of embodiments in conjunction with the accompanying drawings, in which:
[0019] Figure 1 This is a schematic diagram illustrating an application scenario of a method for filtering an image according to an embodiment of the present disclosure;
[0020] Figure 2 A flowchart illustrating a method for filtering an image to be filtered according to an embodiment of the present disclosure is shown;
[0021] Figure 3 A schematic diagram of the bounding box image is shown;
[0022] Figure 4 A schematic diagram of pixel-level blending is shown;
[0023] Figure 5 A schematic diagram of feature layer blending is shown;
[0024] Figure 6 A schematic diagram of an apparatus for filtering an image to be filtered according to an embodiment of the present disclosure is shown. Detailed Implementation
[0025] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.
[0026] It should be understood that the steps described in the method embodiments of this disclosure may be performed in different orders and / or in parallel. Furthermore, method embodiments may include other steps and / or omit certain steps.
[0027] The term "comprising" and its variations as used herein are open-ended inclusions, meaning "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". Definitions of other terms will be given in the description below.
[0028] It should be understood that the concepts of "first" and "second" mentioned in this disclosure are used only to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependencies.
[0029] It should be noted that the terms "a" and "a plurality of" used in this disclosure are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".
[0030] As mentioned above, generative models can generate new data based on existing data to expand / augment the training dataset for downstream tasks (such as the deep learning object detection model described above). However, data generation is difficult to control precisely. For large-scale generated data, it is necessary to filter data that is beneficial to downstream tasks for training, thereby improving the performance of downstream tasks. For example, for the object detection task of diseased leaves, augmented images can be generated using generative models based on real images of diseased leaves. Augmented images with the correct category (i.e., diseased leaf type) and high clarity and realism need to be selected to improve the performance of downstream object detection models. However, the disease features of diseased leaves are usually not obvious, so it is impossible to determine which augmented images are realistic images containing the features of diseased leaves. If all augmented images are used directly for training downstream object detection models without screening, it will have side effects, dragging down the performance of downstream object detection models.
[0031] Currently, it is known that similarity can be calculated one-to-one between real and augmented images, and a similarity threshold can be used to determine whether to keep the augmented image as training data or filter it. The problem with this approach is that it only selects images similar to real images as augmented images, resulting in limited diversity and potentially leading to overfitting, which in turn degrades the performance of downstream tasks.
[0032] This disclosure is made in view of the above-mentioned problems. The purpose of this disclosure is to provide a method, apparatus, and computer program product for filtering images, which can determine the similarity between one or more bounding box images corresponding to one or more bounding box images included in an image to be filtered, based on a reference set of bounding box images, and then determine whether to filter the image to be filtered based on these similarities. Since the reference set of bounding box images includes more bounding box images than the true set of bounding box images, more images to be filtered that are beneficial to downstream tasks can be selected for training of downstream tasks instead of being filtered out, thereby increasing the diversity of the filtered images, reducing overfitting, and thus improving the performance of downstream tasks.
[0033] Figure 1 This is a schematic diagram illustrating an application scenario 100 of a method for filtering an image according to an embodiment of the present disclosure. For example... Figure 1 As shown, the method 102 for filtering an image according to an embodiment of the present disclosure can receive the image to be filtered for filtering. In some embodiments, the image to be filtered may be, for example, a... Figure 1The image generated by the generative model 101, shown in the dashed box, is intended to be used as... Figure 1 The training data for downstream task 103 is shown in the dashed box. The generative model 101 can be, for example, a diffusion model. To improve the performance of downstream task 103, the image to be filtered is filtered using method 102, and the resulting filtered image is input into downstream task 103 as its training data. In some embodiments, the downstream task 103 can be, for example, an object detection model trained to subsequently perform object detection tasks such as diseased leaf detection or face recognition.
[0034] For simplicity, the following description of embodiments of this disclosure primarily uses the detection of diseased leaves as the target detection task. It should be understood that this disclosure is not limited to such application scenarios.
[0035] Figure 2 A flowchart illustrating a method 200 for filtering an image to be filtered according to an embodiment of the present disclosure is shown. In some embodiments, the image to be filtered may be, for example, a... Figure 1 The image generated by the generative model 101 is shown within the dashed box. The generative model 101 can be, for example, a diffusion model. Figure 2 As shown, the method 200 for filtering an image to be filtered according to an embodiment of the present disclosure includes steps S210 to S230.
[0036] In step S210, a reference bounding box image set can be derived based on the set of real bounding box images. Here, the set of real bounding box images is a group of real bounding box images, each of which can be a bounding box image determined from a real image. In this disclosure, the image bounded by the bounding box is described as a bounding box image. As mentioned above, a bounding box is a rectangular region used to determine the location of a target object in an image. The bounding box can be determined manually or automatically using various existing methods, thereby allowing the determination of the bounding box image.
[0037] Figure 3 A schematic diagram of a bounding box image is shown. As an example, such as... Figure 3 Image 300 shown is an image of a leaf, which includes three bounding box images 301, 302, and 303, identified by three rectangles. Taking the target detection task of detecting diseased leaves as an example, the three bounding box images 301, 302, and 303 can be, for example, images labeled / extracted / identified from image 300 that contain yellowed leaves, curled leaves, rotten leaves, withered leaves, or diseased leaves.
[0038] Bounding box images may overlap (e.g., bounding box images 301, 302) or not overlap (e.g., bounding box images 301, 303). Furthermore, it can be understood that although... Figure 3The image shows three bounding box images, but this is only an example, and this disclosure does not limit the number of bounding box images.
[0039] Return to reference Figure 2 As described above, in step S210, a set of reference bounding box images can be derived based on the set of real bounding box images. The set of real bounding box images is a set of real bounding box images, while the set of reference bounding box images is a set of reference bounding box images. The reference bounding box images in the set of reference bounding box images may include (1) real bounding box images from the set of real bounding box images; and (2) augmented bounding box images that are different from the real bounding box images. That is, the set of reference bounding box images includes more reference bounding box images than the set of real bounding box images. By introducing additional augmented bounding box images, embodiments of this disclosure enable more images to be filtered that are beneficial to downstream tasks to be selected for training of downstream tasks instead of being filtered out. For example, for images to be filtered with good diversity, filtering out images that are not similar to real bounding box images can be avoided. Thus, the diversity of filtered images can be increased, overfitting can be reduced, and the performance of downstream tasks can be improved.
[0040] In some embodiments, the known mixup method can be used to augment the ground bounding box image to derive the augmented bounding box image. The mixup method is a data augmentation strategy primarily used in computer vision. Its core idea is to mix input data through simple linear transformations to expand the dataset and enhance the model's generalization ability. Therefore, in some embodiments, step S210 may include deriving an additional bounding box image as the augmented bounding box image based on the ground bounding box images in the ground bounding box image set using the mixup method. That is, the augmented bounding box image can be obtained by mixing multiple ground bounding box images using the mixup method. This allows for the acquisition of features not previously captured in the ground images. For example, in the case of detecting diseased leaves, the ground bounding box image obtained from the ground image may only include images of yellowed and withered leaves, while the augmented bounding box image may include not only images of yellowed and withered leaves, but also images of leaves that are both yellowed and withered.
[0041] The mixup method can involve pixel-level mixing or feature-level mixing. Therefore, in some embodiments, the above-described method of using the mixup method to derive additional bounding box images as augmented bounding box images based on real bounding box images from a set of real bounding box images may include: using the mixup method to perform pixel-level mixing or feature-level mixing on the real bounding box images to derive additional bounding box images as augmented bounding box images. Figure 4 This diagram illustrates pixel-level blending. Pixel-level blending of images can refer to randomly selecting two images and blending them in a certain proportion to generate a new image. For example... Figure 4 As shown, the first bounding box image 401 and the second bounding box image 402 can be pixel-level blended to obtain the third bounding box image 403, which is the additional bounding box image that can be used as an augmented bounding box image. Figure 5 This diagram illustrates feature layer blending. Feature layer blending of an image can refer to blending within randomly selected hidden layers (intermediate layers) of a neural network. For example... Figure 5 As shown, for the first bounding box image 501 and the second bounding box image 502, an encoder and a decoder can be used to randomly select a layer as a mixing layer, and then the features of the layer are linearly interpolated to mix at the feature level. Subsequently, an additional bounding box image is output as an augmented bounding box image based on the mixed features.
[0042] The input data for the mixup blending method described above can include two or more ground truth bounding box images, such as those referenced above. Figure 4 The first bounding box image 401 and the second bounding box image 402 described, or refer to Figure 5 The description includes a first bounding box image 501 and a second bounding box image 502. These ground truth bounding box images can be blended pairwise or two or more ground truth bounding box images can be blended at once. These ground truth bounding box images can be randomly selected from a set of ground truth bounding box images, can be multiple ground truth bounding box images with high similarity from the set of ground truth bounding box images, or can be multiple ground truth bounding box images with large differences from the set of ground truth bounding box images.
[0043] In some embodiments, multiple ground truth bounding box images with high similarity (i.e., small differences) can be selected to derive the augmented bounding box image. By using highly similar ground truth bounding box images for deriving, an augmented bounding box image similar to but slightly different from the ground truth bounding box images can be obtained. This helps reduce the dependence of downstream tasks on certain attributes, thereby improving the generalization ability of downstream tasks. Furthermore, deriving the augmented bounding box image based on multiple highly similar ground truth bounding box images also helps downstream tasks learn more robust feature representations, enabling them to maintain stable output performance even when faced with minor changes in the input image. Therefore, in some embodiments, the above-described method of deriving an additional bounding box image as an augmented bounding box image based on ground truth bounding box images from a set of ground truth bounding box images using the mixup method can include: using the mixup method to mix multiple ground truth bounding box images from the set of ground truth bounding box images with a similarity higher than a third threshold to derive the additional bounding box image as the augmented bounding box image.
[0044] In some embodiments, multiple ground truth bounding box images with low similarity (i.e., large differences) can be selected to derive the augmented bounding box image. Using ground truth bounding box images with low similarity increases the diversity of the reference bounding box image set, which helps downstream tasks learn more generalized feature representations, thereby improving the performance of downstream tasks when faced with new and unseen data. Therefore, in some embodiments, the above-described method of deriving an additional bounding box image as an augmented bounding box image based on ground truth bounding box images from the ground truth bounding box image set using the mixup method may include: using the mixup method to mix multiple ground truth bounding box images from the ground truth bounding box image set whose similarity is below a fourth threshold to derive the additional bounding box image as the augmented bounding box image.
[0045] Return to reference Figure 2Following step S210, the process proceeds to step S220. In step S220, the similarity between each bounding box image in the one or more bounding box images included in the image to be filtered and the reference bounding box image set derived in step S210 can be calculated to obtain one or more bounding box image similarities corresponding to the one or more bounding box images. As described above, bounding boxes can be determined manually or automatically using various existing methods, thereby determining the bounding box images. Therefore, one or more bounding box images included in the image to be filtered can also be determined. In step S220, the similarity between each of these bounding box images and the reference bounding box image set can be determined; that is, for each bounding box image included in the image to be filtered, a similarity value corresponding to that bounding box image is calculated, which characterizes the degree of similarity between the bounding box image and the reference bounding box image set. Thus, one or more bounding box image similarities corresponding to the one or more bounding box images included in the image to be filtered can be obtained. In other words, if the number of bounding box images included in the image to be filtered is N (N is an integer greater than or equal to 1), then in step S220, the similarity of N bounding box images can be calculated, and these N bounding box image similarities correspond one-to-one with the N bounding box images.
[0046] There can be various methods for calculating the similarity between each bounding box image and a set of reference bounding box images. As a simple example, the similarity between a bounding box image and a given bounding box image can be calculated as the maximum similarity among all reference bounding box images in the set of reference bounding box images. Therefore, in some embodiments, step S220 may include, for each bounding box image included in the image to be filtered, calculating the similarity between that bounding box image and each reference bounding box image in the set of reference bounding box images to obtain a plurality of first similarities; and setting the maximum similarity among the plurality of first similarities as the similarity to the bounding box image corresponding to that bounding box image.
[0047] It is understood that any method known in the art for calculating the similarity between two images can be used to calculate the aforementioned first similarity between each bounding box image included in the image to be filtered and each reference bounding box image in the set of reference bounding box images. For example, methods for calculating the similarity between two images may include cosine similarity based on the contrastive language-image pre-training (CLIP) feature vectors of the two images, structural similarity (SSIM), peak signal-to-noise ratio (PSNR), learned perceptual image patch similarity (LPIPS), Euclidean distance, Manhattan distance between the feature vectors of the two images, etc. Assuming that the number of reference bounding box images in the set of reference bounding box images is M (M is an integer greater than or equal to 1), then for each bounding box image, M first similarities can be calculated, and the maximum similarity among these M similarities can be set as the similarity to the bounding box image corresponding to that bounding box image.
[0048] In other embodiments, other methods may be used to calculate the similarity between each bounding box image and the set of reference bounding box images. For example, the similarity between a bounding box image and a reference bounding box image may be calculated as the mean, median, etc., of the similarity between that bounding box image and all reference bounding box images in the set of reference bounding box images. This disclosure does not limit this.
[0049] In step S230, it can be determined whether to filter the image to be filtered based on one or more bounding box image similarities obtained in step S220. In embodiments of this disclosure, the determination of whether to filter the image to be filtered depends on the similarity at the bounding box image level, rather than the similarity at the entire image level. This allows for a finer-grained determination of whether to retain or filter the image to be filtered, making it possible for more images beneficial to downstream tasks to be retained as training data for downstream tasks, thus enhancing the diversity of training data.
[0050] In some embodiments, for an image to be filtered, the image is retained as long as the similarity between any bounding box image in the image to be filtered and the bounding box image corresponding to it is sufficiently high. That is, if the similarity between the bounding box images corresponding to all bounding box images in the image to be filtered is not high enough, the image to be filtered is filtered out. In this case, step S230 may include: filtering out the image to be filtered based on the fact that the similarity between one or more bounding box images obtained in step S220 is less than a first threshold. In this embodiment, the entire image to be filtered is retained as long as it includes a bounding box image that meets the first threshold. Compared with existing schemes that determine whether to filter based on the similarity of the entire image (i.e., the similarity corresponding to each bounding box image needs to meet the threshold), the scheme of this embodiment can allow more diverse images to be retained. Thus, the diversity of filtered images can be increased, thereby improving the performance of downstream tasks.
[0051] Since the reference bounding box images in the reference bounding box image set are already an augmented set compared to the true bounding box image set, in some embodiments, more stringent filtering conditions can be considered for a single image to be filtered. That is, the image to be filtered is filtered out if the similarity of any bounding box image corresponding to it is less than a threshold. In this case, step S230 may include: filtering out the image to be filtered based on at least one of the similarity values of one or more bounding box images obtained in step S220 being less than a second threshold. Since all images to be filtered can be used as training data for downstream tasks after filtering, the above embodiments can reduce outlier data in the training data and improve the accuracy of downstream tasks.
[0052] The first, second, third, and fourth thresholds mentioned above are all similarity thresholds. Their values can be the same or different, and they can be set according to experience or needs.
[0053] In the method for filtering an image according to embodiments of the present disclosure, the diversity of the filtered images can be increased, overfitting can be reduced, and the performance of downstream tasks can be improved.
[0054] Embodiments of this disclosure also provide an apparatus for filtering an image to be filtered. Figure 6 A schematic diagram of a device 600 for filtering an image to be filtered according to an embodiment of the present disclosure is shown.
[0055] like Figure 6 As shown, the device 600 for filtering images according to this embodiment includes a processor 610 and a memory 620. The memory 620 stores one or more computer programs.
[0056] Processor 610 is a program control device such as a microprocessor, which operates according to a program stored in memory 620. Memory 620 is, for example, a storage element such as ROM or RAM. The program executed by processor 610, etc., is stored in memory 620. Figure 6 The apparatus 600 shown for filtering images can be used to implement the method for filtering images disclosed in this application.
[0057] In the device according to embodiments of the present disclosure, the diversity of filtered images can be increased, overfitting can be reduced, and the performance of downstream tasks can be improved.
[0058] Embodiments of this disclosure also provide a computer program product having instructions stored thereon that, when executed by a processor, cause the processor to perform a method for filtering an image to be filtered according to embodiments of this disclosure.
[0059] The hardware computing device or its components described in this disclosure can be implemented by various suitable hardware means, including but not limited to FPGA, ASIC, SoC, discrete gate or transistor logic, discrete hardware components, or any combination thereof. The devices, apparatuses, methods, and systems involved in this disclosure are not limited to any particular hardware architecture or configuration. Components in the disclosed devices, apparatuses, and systems can be discrete or integrated, and can be combined in different ways and / or replaced or supplemented by other components. It should be understood that the teachings of this disclosure can be implemented in various forms, including hardware, software, firmware, dedicated processors, or combinations thereof.
[0060] The block diagrams of the apparatuses, devices, methods, and systems disclosed herein are merely exemplary and are not intended to require or imply that connections, arrangements, or configurations must be made in the manner shown in the block diagrams. As those skilled in the art will recognize, these circuits, devices, apparatuses, and systems can be connected, arranged, and configured in any manner that achieves the desired purpose.
[0061] In the foregoing description, the present invention has been described based on embodiments. These embodiments are merely illustrative, and those skilled in the art should understand that the combination of constituent elements and processes of these embodiments can be modified in various ways, and such modifications are also within the scope of the present invention.
Claims
1. A method for filtering an image to be filtered, comprising: A reference bounding box image set is derived from the set of real bounding box images. The reference bounding box images in the reference bounding box image set include real bounding box images in the set of real bounding box images and augmented bounding box images that are different from the real bounding box images. Calculate the similarity between each bounding box image in one or more bounding box images included in the image to be filtered and the reference bounding box image set, so as to obtain the similarity of one or more bounding box images corresponding to the one or more bounding box images; as well as Whether to filter the image to be filtered is determined based on the similarity of the one or more bounding box images.
2. The method according to claim 1, wherein, Calculate the similarity between each bounding box image and the set of reference bounding box images to obtain the similarity between the bounding box images and the bounding box images, including: Calculate the similarity between the bounding box image and each reference bounding box image in the set of reference bounding box images to obtain multiple first similarities; and The maximum similarity among the plurality of first similarities is set as the similarity to the bounding box image corresponding to the bounding box image.
3. The method according to claim 1, wherein, The step of determining whether to filter the image to be filtered based on the similarity of the one or more bounding box images includes: The images to be filtered are removed based on the fact that the similarity of one or more bounding box images is less than a first threshold.
4. The method according to claim 1, wherein, The step of determining whether to filter the image to be filtered based on the similarity of the one or more bounding box images includes: The images to be filtered are removed based on the fact that at least one of the similarity scores of the one or more bounding box images is less than a second threshold.
5. The method according to claim 1, wherein, The process of deriving a reference bounding box image set based on a set of real bounding box images includes: Based on the real bounding box images in the set of real bounding box images, the mixup method is used to derive additional bounding box images as augmented bounding box images.
6. The method according to claim 5, wherein, The step of using the mixup method to derive additional bounding box images as augmented bounding box images from the set of real bounding box images includes: The Mixup blending method is used to perform pixel-level blending or feature-level blending on the ground bounding box image to derive an additional bounding box image as the augmented bounding box image.
7. The method according to claim 5, wherein, The step of using the mixup method to derive additional bounding box images as augmented bounding box images from the set of real bounding box images includes: The mixup method is used to mix multiple real bounding box images in the set of real bounding box images whose similarity is higher than a third threshold, so as to derive additional bounding box images as augmented bounding box images.
8. The method according to claim 5, wherein, The step of using the mixup method to derive additional bounding box images as augmented bounding box images from the set of real bounding box images includes: The mixup method is used to mix multiple real bounding box images in the set of real bounding box images whose similarity is lower than a fourth threshold, so as to derive additional bounding box images as augmented bounding box images.
9. An apparatus for filtering an image to be filtered, comprising: processor; and Memory, which stores one or more computer programs; Wherein, when the one or more computer programs are run by the processor, the processor causes the processor to perform the method according to any one of claims 1-8.
10. A computer program product having instructions stored thereon, which, when executed by a processor, cause the processor to perform the method according to any one of claims 1-8.