Data augmentation method and device

Abstract

The invention discloses a data augmentation method and device, and relates to the technical field of data augmentation, and the data augmentation method comprises the steps: separating data of a source image according to a target and a scene, and splitting the data into a target image and a scene image; generating a plurality of three-dimensional new view angle target images based on the target image; and fusing the new view target image with the scene image to obtain an expanded data set. According to the invention, through the target new view angle generation algorithm, the new three-dimensional view angle can be generated from the existing data, and the diversity of the data set is greatly enriched.

Description

Technical Field

[0001] The present application relates to the field of data augmentation technology, and in particular to a data augmentation method and device. Background Art

[0002] Data augmentation technology plays a crucial role in machine learning and deep learning, and is widely used in computer vision tasks. In certain domains, data scarcity and data acquisition difficulties are particularly prominent. Flying data collection methods are not only costly but also subject to environmental constraints and security risks, making data acquisition complex and challenging. Furthermore, due to the unique nature and confidentiality of certain data, data sources are extremely limited, making it difficult to meet the requirements for training high-quality models. These issues, including data scarcity, insufficient data diversity, and high data acquisition costs, severely impact model training effectiveness and generalization capabilities.

[0003] Traditional data augmentation methods mainly include geometric transformation, color transformation and image enhancement. These methods generate diverse training samples through operations such as rotation, translation, scaling, shearing, flipping, cropping, and brightness and contrast adjustment, thereby improving the generalization ability of the model.

[0004] Geometric transformation methods, such as rotation and translation, can change the position and perspective of an image through simple two-dimensional transformations. Color transformation and image enhancement techniques, such as brightness adjustment, contrast adjustment, and noise addition, can increase the diversity of the dataset. Furthermore, advanced data augmentation methods such as Cutout, Mixup, and CutMix further increase the complexity of the data through random occlusion and image blending.

[0005] Deep learning-based image generation algorithms are currently gaining traction. Models trained on large amounts of data have demonstrated remarkable performance in generating innovative and diverse images. Currently, widely used text-to-image models generate images based on input text descriptions. These models typically consist of two components: a text encoder and an image generator. The text encoder first converts the input text into a highly understandable vector, while the image generator generates an image based on this vector.

[0006] New perspective generation algorithms leverage large-scale diffusion models to learn geometric priors about natural images. Through various adjustments and training schemes, they can generate high-quality, consistent multi-view images from a single image. Generating new perspectives can be categorized into target perspective generation and scene perspective generation. Scene perspective generation is more challenging and has received less research, while target perspective generation has achieved promising results, paving the way for accurate 3D perspective conversion in data augmentation.

[0007] However, there are few existing studies on applying target new perspective generation to data augmentation, and there are only theoretical studies, lacking feasible strategies for practical applications. Summary of the Invention

[0008] The present application provides a data augmentation method and device, which can generate new three-dimensional perspectives from existing data through a target new perspective generation algorithm, greatly enriching the diversity of the data set.

[0009] In a first aspect, an embodiment of the present application provides a data augmentation method, the data augmentation method comprising:

[0010] Separate the source image data into target image and scene image according to the target and scene;

[0011] generating a plurality of three-dimensional new-viewpoint target images based on the target image;

[0012] The new perspective target image is fused with the scene image to obtain an expanded data set.

[0013] In conjunction with the first aspect, in one embodiment, separating the source image data according to the target and the scene into the target image and the scene image includes:

[0014] The target in the data is masked and image repaired to obtain a scene image after removing the target, so as to split it into a target image and a scene image.

[0015] In combination with the first aspect, in one implementation, fast Fourier convolution and perceptual loss function are used to perform image inpainting.

[0016] In combination with the first aspect, in one embodiment, generating a new-viewpoint target image based on the target image includes:

[0017] Based on the target image, generating a plurality of new perspective target images at predefined angles;

[0018] Multiple new-view target images are spliced ​​together for joint modeling, and the generated images are enhanced based on a linear noise plan and a reference attention mechanism.

[0019] In combination with the first aspect, in one embodiment, fusing the new perspective target image with the scene image includes:

[0020] Determining an area to be fused according to a contour shape of the new-view target image, so as to determine a position of an insertion area on the scene image;

[0021] Solving the Poisson equation through an iterative algorithm to calculate pixel values ​​of the inserted area so that the gradient of the inserted area matches the gradient of the new perspective target image;

[0022] The calculated pixel values ​​are filled into the scene image to complete the fusion of the new perspective target image and the scene image.

[0023] In combination with the first aspect, in one embodiment, the validity of the augmented data obtained by fusion of the new-viewpoint target image and the scene image is verified.

[0024] In conjunction with the first aspect, in one embodiment, verifying the validity of the augmented data obtained by fusing the new-perspective target image with the scene image includes:

[0025] The augmented data is divided into three parts: the left side perspective data T1, the left rear perspective data T2 and the left front perspective data T3 of the target;

[0026] Separate the target and scene from the T1 data, generate a new perspective target image, and fuse the target and scene to generate the target's left rear perspective verification data G1 and left front perspective verification data G2;

[0027] The T1, G1, and G2 data were labeled, and the yolov8 detection model was used to train only the T1 data to obtain M1, and the yolov8 detection model was used to train the T1+G1+G2 data to obtain M2;

[0028] Use M1 and M2 to perform tests on the T2 and T3 test sets in turn to obtain the precision-recall rate (PR) curves of the tests on each test set.

[0029] In a second aspect, an embodiment of the present application provides a data augmentation device, the data augmentation device comprising:

[0030] An image preprocessing module is used to separate the source image data into target images and scene images according to the target and scene.

[0031] A generating module, which generates a plurality of three-dimensional new-viewpoint target images based on the target image;

[0032] A fusion module is used to fuse the new perspective target image with the scene image to obtain an expanded data set.

[0033] In conjunction with the second aspect, in one embodiment, the method further includes:

[0034] A verification module is used to verify the validity of the augmented data after the new perspective target image is fused with the scene image.

[0035] In conjunction with the second aspect, in one embodiment, the verification module verifies the validity of the augmented data obtained by fusing the new-viewpoint target image with the scene image, including:

[0036] The augmented data is divided into three parts: the left side perspective data T1, the left rear perspective data T2 and the left front perspective data T3 of the target;

[0037] Separate the target and scene from the T1 data, generate a new perspective target image, and fuse the target and scene to generate the target's left rear perspective verification data G1 and left front perspective verification data G2;

[0038] The T1, G1, and G2 data were labeled, and the yolov8 detection model was used to train only the T1 data to obtain M1, and the yolov8 detection model was used to train the T1+G1+G2 data to obtain M2;

[0039] Use M1 and M2 to perform tests on the T2 and T3 test sets in turn to obtain the precision-recall rate (PR) curves of the tests on each test set.

[0040] The beneficial effects of the technical solutions provided in the embodiments of the present application include at least:

[0041] The data augmentation method in the present application separates the data of the source image into target images and scene images according to the target and scene; generates multiple three-dimensional new-perspective target images based on the target image; and fuses the new-perspective target images with the scene image to obtain an expanded data set.

[0042] This application, through its algorithm for generating new target perspectives, is able to generate new three-dimensional perspectives from existing data. This breaks the limitation of traditional data augmentation methods, which only perform simple transformations on two-dimensional images, and significantly enriches the diversity of datasets. Furthermore, since it can provide target data from different angles in the same scene, the proposed method can be used to augment data from a single angle to multiple angles, saving both acquisition costs and time, even for data with high acquisition costs, complex acquisition processes, and scarce sources. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 A flowchart of an embodiment of the data augmentation method of this application;

[0044] Figure 2 This is a structural block diagram of an embodiment of the data augmentation device of the present application;

[0045] Figure 3 This is a schematic diagram of the large mask repair model structure for this application;

[0046] Figure 4 This is a schematic diagram of the data splitting results for this application;

[0047] Figure 5 Generate a schematic diagram of the results for a new perspective on the application's goals;

[0048] Figure 6 This is a schematic diagram of the Poisson fusion results of this application;

[0049] Figure 7 This is the flow chart for verifying the validity of this application;

[0050] Figure 8 PR curves of M1 and M2 on the T2 test set for this application;

[0051] Figure 9 These are the PR curves of M1 and M2 in this application on the T3 test set. DETAILED DESCRIPTION

[0052] In order to enable those skilled in the art to better understand the present invention, the following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0053] In order to solve the problems that traditional data augmentation methods cannot provide new three-dimensional perspective information and that deep learning-based text graph models are difficult to generate data that accurately matches existing data, and to break through the limitations of two-dimensional transformation of existing data augmentation methods, based on given data, this paper proposes a data augmentation method that utilizes a target new perspective generation algorithm and Poisson fusion.

[0054] In order to make the objectives, technical solutions and advantages of this application clearer, the implementation methods of this application will be further described in detail below with reference to the accompanying drawings.

[0055] In a first aspect, embodiments of the present application provide a data augmentation method.

[0056] In one embodiment, referring to Figure 1 As shown, Figure 1 This is a flow chart of an embodiment of the data augmentation method of this application. Figure 1 As shown, data augmentation methods include:

[0057] S1, separating the source image data into target image and scene image according to the target and scene;

[0058] S2. generating a plurality of three-dimensional new-viewpoint target images based on the target image;

[0059] S3. Fusing the new perspective target image with the scene image to obtain an expanded data set.

[0060] In addition, after step S3, the validity of the augmented data obtained by fusion of the new-view target image and the scene image is also verified.

[0061] To achieve the above steps, this application includes four functional modules: image preprocessing module, generation module, fusion module and verification module.

[0062] The image preprocessing module splits the existing data into a target image and a scene image. The target image is used for subsequent new perspective generation, while the scene image provides the scene for the newly generated new perspective target image. The generation module uses the zero123++ image conditional diffusion model to stitch together multiple perspective images for joint modeling, ensuring consistency across different perspectives. A linear noise plan is used instead of the original scaled linear plan to enhance the model's global consistency while maintaining its ability to capture detail, enabling the generation of consistent multi-perspective target images from a single-perspective target input. The fusion module utilizes the Poisson algorithm to minimize the gradient between the scene image and the target image. Within the interpolated region, the Poisson fusion algorithm calculates the pixel values ​​within the interpolated region so that their gradients match those of the target image as closely as possible. The pixel values ​​at the boundary are also guaranteed to match those of the target image, ensuring a smooth transition between the interpolated region and the target image. The verification module generates a batch of data with new 3D target perspectives using the target-scene fusion module. This data is then tested using the yolov8 detection algorithm to verify the effectiveness of the proposed data augmentation method.

[0063] This application introduces the new perspective generation technology in the field of image generation into the field of data augmentation, realizing data expansion from a higher level dimension. Figure 2 As shown, the collected data is preprocessed to separate the data into objects and scenes. The target data is then subjected to the latest new perspective generation technology to generate data from different perspectives. The generated multi-perspective object data is then combined with the scene to create an augmented dataset. Compared to the original data, the generated data retains the same scene but provides new 3D perspectives of the objects, which is particularly important for tasks requiring multi-perspective data. It is important to note that the existing zero123++ image diffusion model is a variant of the Stable Diffusion model, preserving as much of the prior knowledge from the Stable Diffusion model as possible. Therefore, it can generate a wide range of object types and has extremely high application value.

[0064] The above four functional modules are described in further detail below.

[0065] (1) Image preprocessing module

[0066] like Figure 3As shown in the figure, a data splitting technology is proposed. First, the target in the data is masked, and then the LaMa (Large Mask inpainting) method is used to inpaint the image. Fast Fourier convolutions and high receptive field perceptual loss are used to effectively deal with the problem of inpainting large missing areas and high-resolution images, and the scene data after the target is removed is obtained. Fast Fourier convolution uses Fourier transform to divide the convolution operation into local branches and global branches, so that the network can obtain the global receptive field at an early level. The high receptive field perceptual loss based on the semantic segmentation network is used to ensure that the global structure consistency of the image is maintained during the inpainting process. Compared with the traditional pixel-level loss, this loss function can better handle the ambiguity problem in large mask inpainting. The acquisition of target data is also based on a deep learning model. The existing background removal algorithm library (rembg library) is used to accurately distinguish between foreground and background parts and effectively remove the background. The processed image only contains foreground targets, providing more accurate input data for subsequent data enhancement and target detection. The result of data splitting is shown in the figure Figure 4 shown.

[0067] (2) Generate module

[0068] The generation module inputs the target data with the background removed and processes it through the zero123++ model to generate a series of multi-view images with predefined angles. The improved noise plan and reference attention mechanism enhance the global consistency and detail fidelity of the image generation, ensuring that the output multi-view images are highly consistent and natural in vision. The generated results are as follows: Figure 5 As shown in the figure, the generation module stitches together multiple viewpoint images for joint modeling to ensure the consistency of the generated images across different viewpoints. A linear noise plan is used instead of the original scaled linear plan to enhance the model's performance in terms of global consistency while maintaining its ability to capture details. A scaled reference attention mechanism is introduced to provide more accurate local image conditions, making the generated images more consistent with the characteristics of the input image. Using a method similar to FlexDiffuse, the global image embedding is added to the text embedding to enhance the model's understanding of global semantics and improve the generation quality of unseen areas.

[0069] (3) Fusion module

[0070] The fusion module is mainly implemented through the Poisson Image Editing algorithm, which is an image processing technology that is usually used to seamlessly fuse areas between images. It solves the Poisson equation so that the target image area inserted on the scene image transitions naturally with the surrounding environment, thereby achieving a seamless fusion effect. The main steps of this algorithm include: selecting the target image and the scene image, and defining the area to be fused, determining the position of the inserted area based on the contour shape of the generated new perspective target image, and determining the position of the inserted area on the scene image. The Poisson equation is solved through an iterative algorithm, and the pixel value of the inserted area is calculated so that its gradient matches the gradient of the target image as much as possible. The calculated pixel value is filled into the scene image to complete the seamless fusion of the new perspective target data and the scene data. The fusion result is as follows: Figure 6 In addition, by appropriately scaling the target data of the new perspective during Poisson fusion, low-resolution long-distance small target data is synthesized.

[0071] (4) Verification module

[0072] like Figure 7 As shown in the figure, the verification module divides the real data into three parts: the target's left side perspective data (denoted as T1), the left rear perspective data (T2), and the left front perspective data (T3). Then, the T1 data is used as the input of the above steps, and after image preprocessing, target new perspective generation, and target scene fusion, the target's left rear perspective verification data (denoted as G1) and left front perspective verification data (G2) are generated.

[0073] The T1, G1, and G2 data are labeled, and the yolov8 detection model is used to train only the T1 data to obtain M1. The yolov8 detection model is used to train the T1+G1+G2 data to obtain M2. Then, M1 and M2 are used to detect on the T2 and T3 data sets respectively.

[0074] In the 100 data of T2, the precision of M1 detection bounding box is 0.924, the recall rate is 0.732, the mAP50 is 0.88, and the mAP50-95 is 0.515; the precision of M2 detection bounding box is 0.988, the recall rate is 0.847, the mAP50 is 0.96, and the mAP50-95 is 0.625. Figure 8 are the PR curves of M1 and M2 detected on the T2 test set respectively.

[0075] In the 100 data of T3, the precision of M1 detection bounding box is 0.977, the recall rate is 0.837, the mAP50 is 0.914, and the mAP50-95 is 0.539; the precision of M2 detection bounding box is 0.928, the recall rate is 0.98, the mAP50 is 0.989, and the mAP50-95 is 0.564. Figure 9 The following are the PR curves of M1 and M2 tested on the T3 test set. The experiment proves that the data augmentation method proposed in this application can significantly improve the generalization ability of the small downstream model on the end by expanding the data.

[0076] In summary, the data augmentation method in the present application separates the data of the source image into target images and scene images according to the target and scene; generates multiple three-dimensional new-perspective target images based on the target image; and fuses the new-perspective target images with the scene image to obtain an expanded data set.

[0077] This application, through its algorithm for generating new target perspectives, is able to generate new three-dimensional perspectives from existing data. This breaks the limitation of traditional data augmentation methods, which only perform simple transformations on two-dimensional images, and significantly enriches the diversity of datasets. Furthermore, since it can provide target data from different angles in the same scene, the proposed method can be used to augment data from a single angle to multiple angles, saving both acquisition costs and time, even for data with high acquisition costs, complex acquisition processes, and scarce sources.

[0078] In a second aspect, an embodiment of the present application provides a data augmentation device.

[0079] In one embodiment, referring to Figure 2 As shown, Figure 2 This is a structural block diagram of an embodiment of the data augmentation device of this application. Figure 2 As shown, the data augmentation device includes: an image preprocessing module, a generation module and a fusion module.

[0080] Among them, the image preprocessing module is used to separate the data of the source image according to the target and scene, and split it into target image and scene image; the generation module generates multiple three-dimensional new-perspective target images based on the target image; and the fusion module is used to fuse the new-perspective target image with the scene image to obtain an expanded data set.

[0081] Furthermore, in one embodiment, the present invention further includes:

[0082] A verification module is used to verify the validity of the augmented data after the new perspective target image is fused with the scene image.

[0083] Furthermore, in one embodiment, the verification module verifies the validity of the augmented data obtained by merging the new-viewpoint target image with the scene image, including:

[0084] The augmented data is divided into three parts: the left side perspective data T1, the left rear perspective data T2 and the left front perspective data T3 of the target;

[0085] Separate the target and scene from the T1 data, generate a new perspective target image, and fuse the target and scene to generate the target's left rear perspective verification data G1 and left front perspective verification data G2;

[0086] The T1, G1, and G2 data were labeled, and the yolov8 detection model was used to train only the T1 data to obtain M1, and the yolov8 detection model was used to train the T1+G1+G2 data to obtain M2;

[0087] Use M1 and M2 to perform tests on the T2 and T3 test sets in turn to obtain the precision-recall rate (PR) curves of the tests on each test set.

[0088] Furthermore, in one embodiment, the image preprocessing module separates the source image data according to the target and the scene, and splits the data into the target image and the scene image, including:

[0089] The target in the data is masked and image repaired to obtain a scene image after removing the target, so as to split it into a target image and a scene image.

[0090] Furthermore, in one embodiment, the image preprocessing module uses fast Fourier convolution and perceptual loss function to perform image inpainting.

[0091] Furthermore, in one embodiment, the generating module generates a new-viewing angle target image based on the target image, including:

[0092] Based on the target image, generating a plurality of new perspective target images at predefined angles;

[0093] Multiple new-view target images are spliced ​​together for joint modeling, and the generated images are enhanced based on a linear noise plan and a reference attention mechanism.

[0094] Furthermore, in one embodiment, the fusion module fuses the new-viewpoint target image with the scene image, including:

[0095] Determining an area to be fused according to a contour shape of the new-view target image, so as to determine a position of an insertion area on the scene image;

[0096] Solving the Poisson equation through an iterative algorithm to calculate pixel values ​​of the inserted area so that the gradient of the inserted area matches the gradient of the new perspective target image;

[0097] The calculated pixel values ​​are filled into the scene image to complete the fusion of the new perspective target image and the scene image.

[0098] It should be noted that the serial numbers of the above-mentioned embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0099] The terms "including" and "having" and any variations thereof in the specification and claims of this application and the above-mentioned drawings are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but optionally includes steps or units that are not listed, or optionally includes other steps or units inherent to these processes, methods, products or devices. The terms "first", "second" and "third" are used to distinguish different objects, etc., and do not represent a sequence, nor do they limit the "first", "second" and "third" to different types.

[0100] In the description of the embodiments of this application, the words "exemplary," "for example," or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary," "for example," or "for example" in the embodiments of this application should not be construed as being preferred or advantageous over other embodiments or designs. Rather, the use of words such as "exemplary," "for example," or "for example" is intended to present the relevant concepts in a concrete manner.

[0101] In the description of the embodiments of the present application, unless otherwise specified, “ / ” means or, for example, A / B can mean A or B; “and / or” in the text is merely a description of the association relationship of associated objects, indicating that three relationships may exist, for example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. In addition, in the description of the embodiments of the present application, “multiple” refers to two or more than two.

[0102] In some processes described in the embodiments of the present application, multiple operations or steps are included that appear in a specific order. However, it should be understood that these operations or steps may not be performed in the order in which they appear in the embodiments of the present application or may be performed in parallel. The sequence numbers of the operations are only used to distinguish between different operations, and the sequence numbers themselves do not represent any order of execution. In addition, these processes may include more or fewer operations, and these operations or steps may be performed in sequence or in parallel, and these operations or steps may be combined.

[0103] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus the necessary general hardware platform, of course, it can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes a number of instructions for enabling a terminal device to execute the methods described in each embodiment of the present application.

[0104] The above are only preferred embodiments of the present application and do not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A data augmentation method, characterized in that: The data augmentation method comprises: Separate the source image data into target image and scene image according to the target and scene; generating a plurality of three-dimensional new-viewpoint target images based on the target image; The new perspective target image is fused with the scene image to obtain an expanded data set.

2. The data augmentation method according to claim 1, wherein: The method of separating the source image data into target images and scene images according to the target and the scene includes: The target in the data is masked and image repaired to obtain a scene image after removing the target, so as to split it into a target image and a scene image.

3. The data augmentation method according to claim 2, wherein: Image inpainting using fast Fourier convolution and perceptual loss function.

4. The data augmentation method according to claim 1, wherein: Generating a new-viewing angle target image based on the target image includes: Based on the target image, generating a plurality of new perspective target images at predefined angles; Multiple new-view target images are spliced ​​together for joint modeling, and the generated images are enhanced based on a linear noise plan and a reference attention mechanism.

5. The data augmentation method according to claim 1, wherein: The fusing the new-view target image with the scene image comprises: Determining an area to be fused according to a contour shape of the new-view target image, so as to determine a position of an insertion area on the scene image; Solving the Poisson equation through an iterative algorithm to calculate pixel values ​​of the inserted area so that the gradient of the inserted area matches the gradient of the new perspective target image; The calculated pixel values ​​are filled into the scene image to complete the fusion of the new perspective target image and the scene image.

6. The data augmentation method according to claim 1, wherein: Also includes: Verify the validity of the augmented data after the new perspective target image is fused with the scene image.

7. The data augmentation method according to claim 6, wherein: The verifying the validity of the augmented data after the new perspective target image is fused with the scene image includes: The augmented data is divided into three parts: the left side perspective data T1, the left rear perspective data T2 and the left front perspective data T3 of the target; Separate the target and scene from the T1 data, generate a new perspective target image, and fuse the target and scene to generate the target's left rear perspective verification data G1 and left front perspective verification data G2; The T1, G1, and G2 data were labeled, and the yolov8 detection model was used to train only the T1 data to obtain M1, and the yolov8 detection model was used to train the T1+G1+G2 data to obtain M2; Use M1 and M2 to perform tests on the T2 and T3 test sets in turn to obtain the precision-recall rate (PR) curves of the tests on each test set.

8. A data augmentation device, characterized in that: The data augmentation device comprises: An image preprocessing module is used to separate the source image data into target images and scene images according to the target and scene; A generating module, which generates a plurality of three-dimensional new-viewpoint target images based on the target image; A fusion module is used to fuse the new perspective target image with the scene image to obtain an expanded data set.

9. The data augmentation device according to claim 8, wherein: Also includes: A verification module is used to verify the validity of the augmented data after the new perspective target image is fused with the scene image.

10. The data augmentation device according to claim 9, wherein: The verification module verifies the validity of the augmented data obtained by fusing the new-view target image with the scene image, including: The augmented data is divided into three parts: the left side perspective data T1, the left rear perspective data T2 and the left front perspective data T3 of the target; Separate the target and scene from the T1 data, generate a new perspective target image, and fuse the target and scene to generate the target's left rear perspective verification data G1 and left front perspective verification data G2; The T1, G1, and G2 data were labeled, and the yolov8 detection model was used to train only the T1 data to obtain M1, and the yolov8 detection model was used to train the T1+G1+G2 data to obtain M2; Use M1 and M2 to perform tests on the T2 and T3 test sets in turn to obtain the precision-recall rate (PR) curves of the tests on each test set.

Images