A method, device, medium and electronic equipment for detecting a target at a gigapixel level

Abstract

The present disclosure provides a method and device for detecting a target in a gigapixel image, a medium and an electronic device. The method comprises: detecting a gigapixel image using a pre-trained fast detector to determine a target size distribution of the gigapixel image; removing the background of the gigapixel image using a pre-trained SAM model to determine a background-free gigapixel image region; and detecting the target in the background-free gigapixel image region using the pre-trained fast detector and a pre-trained slow detector according to the target size distribution of the gigapixel image to determine a target detection result. The present disclosure can accurately and efficiently detect the target in the gigapixel image, reduce the false detection and missed detection, reduce the computational complexity, improve the processing speed, and has generalization in large-scale high-pixel image analysis tasks.

Description

Technical Field

[0001] This disclosure relates to the field of image processing technology, and more specifically, to a method, apparatus, medium, and electronic device for detecting billion-pixel-level targets. Background Technology

[0002] Object detection is a core problem in computer vision, aiming to determine the location and size of specific objects in an image. Object detection in gigapixel imaging is one of the most challenging and significant problems within this field. Billapixel object detection offers extremely high resolution, enabling the capture of more, smaller, and finer-grained targets in large-scale scenes, such as urban or satellite imagery. This high-resolution detection capability is crucial for fields such as military reconnaissance, environmental monitoring, and precision agriculture.

[0003] PANDA [Xueyang Wang, Xiya Zhang, Yinheng Zhu, Yuchen Guo, Xiaoyun Yuan, Liuyu Xiang, Zerun Wang, Guiguang Ding, David Brady, Qionghai Dai, et al., “Panda: A gigapixel-level human-centric video dataset,” in CVPR, 2020.] introduced the first-ever gigapixel-level human-centric video dataset. This dataset is specifically designed for large-scale, long-term, and multi-object visual analysis. In recent years, object detection problems for gigapixel imaging have mostly addressed the issue by dividing the images into manageable patches. PAN [Jiahao Fan, Huabin Liu, Wenjie Yang, JohnSee, Aixin Zhang, and Weiyao Lin, “Speed ​​up object detection on gigapixel-level images with patch arrangement,” in CVPR, 2022] introduced an innovative framework that arranges patches into a compact canvas for faster object detection. GigaDet [Kai Chen, Zerun Wang, Xueyang Wang, Dahan Gong, Longlong Yu, Yuchen Guo, and Guiguang Ding, “Towards real-time object detection in gigapixel-level video,” Neurocomputing, 2022] emphasizes a balance between speed and accuracy, employing a unified patch selection mechanism. Meanwhile, Remix [Shiqi Jiang, Zhiqi Lin, Yuanchun Li, Yuanchao Shu, and Yunxin Liu, “Flexible high-resolution object detection on edge devices with tunable latency,” in MobiCom, 2021] processes ultra-high-resolution images on edge devices, adjusting detection based on crowd size and time constraints, but this impacts system stability.

[0004] While the above methods can effectively detect images at the billion-pixel level, they are limited by their massive computational requirements, the various scales of objects they can process, and the surge in background detail, making it impossible to balance accuracy and speed. Summary of the Invention

[0005] In view of the deficiencies in the prior art, the purpose of this disclosure is to provide a method, apparatus, medium and electronic device for detecting billion-pixel-level targets.

[0006] To achieve the above objectives, according to a first aspect of this disclosure, a billion-pixel-level target detection method is provided, comprising:

[0007] A pre-trained fast detector is used to detect and process billion-pixel level images to determine the target size distribution of the billion-pixel level images;

[0008] A pre-trained SAM model is used to remove the background from the billion-pixel image to determine the billion-pixel image region without background.

[0009] Based on the target size distribution of the billion-pixel-level image, the pre-trained fast detector and the pre-trained slow detector are used to perform target detection processing on the background-free billion-pixel-level image region to determine the target detection result.

[0010] Optionally, the step of using a pre-trained fast detector to perform detection processing on the billion-pixel-level image and determining the target size distribution of the billion-pixel-level image includes:

[0011] The pre-trained fast detector is used to perform a first downsampling process on the billion-pixel-level image to determine the first downsampled image;

[0012] The pre-trained fast detector is used to perform target detection processing on the billion-pixel image to determine the target size distribution of the billion-pixel image.

[0013] Optionally, the step of using a pre-trained SAM model to perform background removal processing on the billion-pixel-level image to determine the background-free billion-pixel-level image region includes:

[0014] The pre-trained SAM model is used to perform a second downsampling process on the billion-pixel image to determine the second downsampled image;

[0015] Based on the preset target threshold, the pre-trained SAM model is used to perform region filtering on the second downsampled image to determine the background-free billion-pixel-level image region.

[0016] Optionally, the step of using the pre-trained SAM model to perform region filtering processing on the second downsampled image according to a preset target threshold to determine the background-free billion-pixel-level image region includes:

[0017] Pixels in the second downsampled image that are larger than the preset target threshold are identified as background regions and removed.

[0018] Pixels in the second downsampled image that are smaller than the preset target threshold are identified as the background-free billion-pixel-level image region and retained.

[0019] Optionally, the step of performing target detection processing on the background-free billion-pixel image region using the pre-trained fast detector and the pre-trained slow detector based on the target size distribution of the billion-pixel image, and determining the target detection result, includes:

[0020] The pre-trained fast detector is optimized using TensorRT to determine the optimized fast detector.

[0021] The pre-trained slow detector is converted to half-floating-point precision to determine the converted slow detector.

[0022] Based on the preset detection threshold and the target size distribution of the billion-pixel image, the optimized fast detector and the converted slow detector are used to perform target detection processing on the background-free billion-pixel image region to determine the target detection result.

[0023] Optionally, the step of performing target detection processing on the background-free billion-pixel image region using the optimized fast detector and the converted slow detector based on a preset detection threshold and the target size distribution of the billion-pixel-level image, and determining the target detection result, includes:

[0024] If the target size of the target in the backgroundless billion-pixel-level image region is larger than the preset detection threshold, the optimized fast detector is used to perform target detection processing to determine the first target detection result.

[0025] If the target size of the target in the backgroundless billion-pixel-level image region is not greater than the preset detection threshold, the target detection process is performed using the converted slow detector to determine the second target detection result.

[0026] Optionally, the method further includes:

[0027] Preprocess the pre-defined dataset to determine the training dataset;

[0028] The training dataset is used to train a preset fast detector, a preset slow detector, and a preset SAM model to determine the pre-trained fast detector, the pre-trained slow detector, and the pre-trained SAM model.

[0029] According to a second aspect of this disclosure, a billion-pixel-level target detection device is provided, comprising:

[0030] The first detection and processing module is used to perform detection processing on the billion-pixel-level image using a pre-trained fast detector to determine the target size distribution of the billion-pixel-level image.

[0031] The background processing module is used to perform background removal processing on the billion-pixel-level image using a pre-trained SAM model to determine the billion-pixel-level image region without background.

[0032] The second detection processing module is used to perform target detection processing on the background-free billion-pixel image region using the pre-trained fast detector and the pre-trained slow detector according to the target size distribution of the billion-pixel image, and to determine the target detection result.

[0033] According to a third aspect of this disclosure, a non-transitory computer-readable storage medium is provided, on which a computer program is stored, characterized in that, when executed by a processor, the program implements the steps of the method provided in the first aspect of this disclosure.

[0034] According to a fourth aspect of this disclosure, an electronic device is provided, comprising:

[0035] A memory on which computer programs are stored;

[0036] A processor for executing the computer program in the memory to implement the steps of the method provided in the first aspect of this disclosure.

[0037] Compared with the prior art, the embodiments disclosed herein have at least one of the following beneficial effects:

[0038] The above technical solution employs a pre-trained fast detector to detect and process gigapixel-level images, determining the target size distribution within these images. This pre-trained fast detector performs initial screening, reducing the number of pixels requiring processing, thus lowering computational complexity and increasing processing speed. A pre-trained SAM model is then used for background removal, identifying background-free gigapixel-level image regions, further reducing the target detection range and improving detection speed. Based on the target size distribution of the gigapixel-level images, both pre-trained fast and slow detectors are used to detect targets in these background-free regions, determining the detection results. This approach enables accurate and efficient target detection in complex gigapixel-level images, reducing false positives and false negatives. It has broad application prospects and can be applied to large-scale high-pixel image analysis tasks. Attached Figure Description

[0039] Other features, objects, and advantages of this disclosure will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0040] Figure 1 This is a flowchart illustrating a billion-pixel-level target detection method according to an exemplary embodiment.

[0041] Figure 2 This is a flowchart illustrating a billion-pixel-level target detection method according to another exemplary embodiment.

[0042] Figure 3 This is an image illustration of a billion-pixel-level target detection according to an exemplary embodiment.

[0043] Figure 4 This is a block diagram illustrating a billion-pixel-level target detection device according to an exemplary embodiment. Detailed Implementation

[0044] The present disclosure will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present disclosure, but do not limit the present disclosure in any way. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present disclosure. These all fall within the protection scope of the present disclosure.

[0045] Figure 1 This is a flowchart illustrating a billion-pixel-level target detection method according to an exemplary embodiment.

[0046] like Figure 1 As shown, this disclosure provides a billion-pixel-level target detection method, including steps S11 to S13.

[0047] S11 uses a pre-trained fast detector to perform detection processing on billion-pixel level images to determine the target size distribution of billion-pixel level images.

[0048] The fast detector can employ a YOLO series network model. For example, the initial fast detector disclosed herein employs an initial YOLOv8 network model and pre-trains the initial YOLOv8 network model.

[0049] A pre-trained fast detector is used to detect and process billion-pixel images. Specifically, the YOLOv8 network model is used to perform scale estimation processing on billion-pixel images to determine the target size distribution of each region in the billion-pixel image.

[0050] S12 uses a pre-trained SAM model to perform background removal processing on billion-pixel-level images to determine the billion-pixel-level image regions without background.

[0051] The pre-trained SAM model is the Segment Anything Model.

[0052] S13. Based on the target size distribution of the billion-pixel-level image, pre-trained fast detectors and pre-trained slow detectors are used to perform target detection processing on the billion-pixel-level image region without background, and the target detection result is determined.

[0053] The multi-agent strategy is implemented by using pre-trained fast detectors and pre-trained slow detectors. The fast detector has a faster detection speed, while the slow detector has higher accuracy but a slower detection speed. Thus, the slow detector can be used in smaller target areas, while the fast detector can be used in larger target areas. Both fast and slow detectors are used to detect targets in different target areas, while background areas are not detected.

[0054] The slow detector disclosed herein can use the DINO network model, and the fast detector uses the YOLOv8 network model.

[0055] The above technical solution employs a pre-trained fast detector to detect and process gigapixel-level images, determining the target size distribution within these images. This pre-trained fast detector performs initial screening, reducing the number of pixels requiring processing, thus lowering computational complexity and increasing processing speed. A pre-trained SAM model is then used for background removal, identifying background-free gigapixel-level image regions, further reducing the target detection range and improving detection speed. Based on the target size distribution of the gigapixel-level images, both pre-trained fast and slow detectors are used to detect targets in these background-free regions, determining the detection results. This approach enables accurate and efficient target detection in complex gigapixel-level images, reducing false positives and false negatives. It has broad application prospects and can be applied to large-scale high-pixel image analysis tasks.

[0056] Figure 2 This is a flowchart illustrating a billion-pixel-level target detection method according to another exemplary embodiment.

[0057] like Figure 2 As shown, in some possible embodiments, a billion-pixel-level target detection method further includes S14 to S15.

[0058] S14: Preprocess the preset dataset to determine the training dataset.

[0059] The default dataset is the PANDA dataset.

[0060] Preprocessing of the predefined dataset includes:

[0061] The training data in the preset dataset is segmented using a field of view of 1500, 3000, 6000, and 12000 pixels to improve the size diversity of targets in the training data, thereby enhancing the model's ability to generalize to targets at various scales.

[0062] Add the CrowdHuman, MOT, CUHK-SYSU, and PRW datasets to the preset dataset to enrich the scene information of the training dataset.

[0063] S15, use the training dataset to train the preset fast detector, preset slow detector and preset SAM model to determine the pre-trained fast detector, pre-trained slow detector and pre-trained SAM model.

[0064] As an example, the pre-processed PANDA dataset, i.e. the training dataset, is used to train the initial YOLOv8 network model, the initial DINO network model, and the initial SAM model to determine the pre-trained YOLOv8 network model, i.e. the pre-trained fast detector, the pre-trained DINO network model, i.e. the pre-trained slow detector, and the pre-trained SAM model.

[0065] By using the above technical solution, the pre-processed training dataset is used to train the preset fast detector, the preset slow detector, and the preset SAM model, which can improve the generalization ability of each model to targets at various scales and adapt to large-scale high-pixel image analysis tasks.

[0066] Figure 3 This is an image illustration of a billion-pixel-level target detection according to an exemplary embodiment.

[0067] In some possible embodiments, S11 includes S21 to S22.

[0068] S21, a pre-trained fast detector is used to downsample the billion-pixel image to determine the first downsampled image.

[0069] One approach is to use a pre-trained YOLOv8 network model to downsample the billion-pixel image, and the first downsampled image is denoted as I.

[0070] S22 uses a pre-trained fast detector to perform target detection processing on billion-pixel level images to determine the target size distribution of billion-pixel level images.

[0071] For example, a pre-trained YOLOv8 network model is used to perform object detection processing on a billion-pixel image:

[0072] The first step is to generate a canvas C representing the distribution of the target region, wherein the size of the canvas C is the same as that of the first downsampled image I;

[0073] The second step is to represent the set of targets with detection boxes detected by the pre-trained YOLOv8 network model as D, where each target d∈D;

[0074] The third step is to determine the area A of the detected target. d The area corresponding to each detected target is plotted on the canvas C at the same location as the first downsampled image I, while the area on canvas C containing undetected targets remains zero.

[0075] The fourth step is to determine the average area of ​​region R in the canvas C to effectively filter out regions containing undetected targets and determine a robust size estimate, i.e., a stable and reliable target size distribution for billion-pixel-level images.

[0076]

[0077] Among them, C i This represents the area corresponding to position i on canvas C, and count represents the number of non-zero regions within region R.

[0078] like Figure 3 As shown, Figure 3 The image on the left is a schematic diagram of how a pre-trained fast detector determines the target size distribution of a billion-pixel image.

[0079] By employing the above technical solution, through downsampling processing and preliminary screening by a fast detector, the regions containing occluded or failed-to-detect targets are filtered out, reducing the number of pixels that need to be processed, thereby reducing computational complexity and improving processing speed.

[0080] In some possible embodiments, S12 includes S31 to S32.

[0081] S31, a pre-trained SAM model is used to perform a second downsampling process on the billion-pixel-level image to determine the second downsampled image.

[0082] Among them, a pre-trained SAM model is used to perform a second downsampling process on the billion-pixel-level image to the standard resolution. The second downsampled image can be represented as I1.

[0083] S32, based on the preset target threshold, uses a pre-trained SAM model to perform region filtering on the second downsampled image to determine the billion-pixel-level image region without background.

[0084] The preset target threshold can be the area A of the largest target region in the target size distribution of a billion-pixel image. max .

[0085] In one possible embodiment, a pre-trained SAM model can be used to perform initial segmentation processing on the second downsampled image to determine preliminary segmentation lines, denoted as S. initial :S initial =SAM(I1I).

[0086] Based on the initial dividing line S initial Construct subset S filter :S filter ={s∈S initial |Area(s)>A max}

[0087] As an example, pixels in the second downsampled image that are larger than a preset target threshold are identified as background regions and removed.

[0088] In this disclosure, the subset S to be constructed filter Use it as a background area and remove it.

[0089] As another example, the second step involves using pixels in the image that are smaller than a preset target threshold to identify and preserve background-free billion-pixel-level image regions.

[0090] like Figure 3 As shown, Figure 3 The image in the middle is a schematic diagram of the billion-pixel-level image region without background determined using a pre-trained SAM model.

[0091] The above technical solutions can reduce the number of pixels that need to be processed, narrow the target detection range, and reduce the computational load.

[0092] In some possible embodiments, S13 may include S41 to S43.

[0093] S41, TensorRT is used to optimize the pre-trained fast detector to determine the optimized fast detector.

[0094] Among them, deploying a pre-trained fast detector using TensorRT can improve the target detection speed of the fast detector.

[0095] S42 converts the pre-trained slow detector to half-floating-point precision and determines the converted slow detector.

[0096] Converting the slow detector to half-floating-point precision can improve the detection accuracy of the slow detector.

[0097] S43, based on the preset detection threshold and the target size distribution of the billion-pixel-level image, the optimized fast detector and the converted slow detector are used to perform target detection processing on the billion-pixel-level image region without background, and the target detection result is determined.

[0098] The preset detection threshold can be expressed as T. scale .

[0099] Constructing a multi-agent strategy for fast and slow detectors includes:

[0100]

[0101] Where Detector represents the policy, and Scale regionIndicates the target size.

[0102] In one possible embodiment, if the target size of a target in a billion-pixel-level image region without background is greater than a preset detection threshold, a fast detector with optimized processing is used to perform target detection processing to determine the first target detection result.

[0103] The first target detection result indicates the position and size of a larger target that is greater than the preset detection threshold.

[0104] In another possible embodiment, if the target size of the target in the billion-pixel-level image region without background is not greater than a preset detection threshold, a slow detector that has undergone conversion processing is used to perform target detection processing to determine the second target detection result.

[0105] The second target detection result indicates the position and size of smaller targets that are less than or equal to a preset detection threshold.

[0106] like Figure 3 As shown, Figure 3 The image on the right is a schematic diagram of object detection processing using a pre-trained fast detector and a pre-trained slow detector.

[0107] In one possible embodiment, based on the unified hardware conditions of the NVIDIA Jetson AGX Orin 32G, the billion-pixel-level object detection method disclosed herein, along with hsslab_inspur, AIOT LAB, Mirror_Descent, and libra methods, were validated on the PANDA dataset, namely the CACAI Gigavision challenge. Table 1 below shows the validation results of the above five methods.

[0108]

[0109] Table 1

[0110] As shown in Table 1, the billion-pixel-level object detection method provided in this disclosure ranks first in the CACAI Gigavision Challenge. This billion-pixel-level object detection method employs a three-pronged strategy of "estimation, screening, and delegation," utilizing free-scale estimation, a large-scale foundational model, and a specialized detector to create a finely tuned trajectory for large-scale high-pixel analysis. This maintains robustness and accuracy in large-scale image analysis tasks, improving the accuracy and efficiency of object detection.

[0111] Figure 4 This is a block diagram illustrating a billion-pixel-level target detection device according to an exemplary embodiment.

[0112] Based on the same concept, this disclosure also provides a billion-pixel-level target detection device 100, such as Figure 4 As shown, it includes: a first detection processing module 110, a background processing module 120, and a second detection processing module 130.

[0113] The first detection and processing module 110 is used to perform detection processing on the billion-pixel-level image using a pre-trained fast detector to determine the target size distribution of the billion-pixel-level image.

[0114] Background processing module 120 is used to perform background removal processing on billion-pixel level images using a pre-trained SAM model to determine the billion-pixel level image region without background.

[0115] The second detection processing module 130 is used to perform target detection processing on the backgroundless billion-pixel image region using a pre-trained fast detector and a pre-trained slow detector based on the target size distribution of the billion-pixel image, and to determine the target detection result.

[0116] The above technical solution employs a pre-trained fast detector to detect and process gigapixel-level images, determining the target size distribution within these images. This pre-trained fast detector performs initial screening, reducing the number of pixels requiring processing, thus lowering computational complexity and increasing processing speed. A pre-trained SAM model is then used for background removal, identifying background-free gigapixel-level image regions, further reducing the target detection range and improving detection speed. Based on the target size distribution of the gigapixel-level images, both pre-trained fast and slow detectors are used to detect targets in these background-free regions, determining the detection results. This approach enables accurate and efficient target detection in complex gigapixel-level images, reducing false positives and false negatives. It has broad application prospects and can be applied to large-scale high-pixel image analysis tasks.

[0117] Regarding the embodiments of the above system, the specific ways in which each module performs operations have been described in detail in the embodiments of the method, and will not be elaborated here.

[0118] Based on the same concept described above, in another embodiment of this disclosure, an electronic device is also provided, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes a billion-pixel-level target detection method for execution.

[0119] Optionally, the memory is used to store programs; the memory may include volatile memory, such as random-access memory (RAM), such as static random-access memory (SRAM), double data rate synchronous dynamic random-access memory (DDR SDRAM), etc.; the memory may also include non-volatile memory, such as flash memory. The memory is used to store computer programs (such as application programs, functional modules, etc. that implement the above methods), computer instructions, etc., and the aforementioned computer programs, computer instructions, etc., can be partitioned and stored in one or more memories. Furthermore, the aforementioned computer programs, computer instructions, data, etc., can be accessed by the processor.

[0120] The aforementioned computer programs, computer instructions, etc., can be stored in partitions within one or more memory locations. Furthermore, the aforementioned computer programs, computer instructions, data, etc., can be accessed by a processor.

[0121] A processor is used to execute a computer program stored in memory to implement the various steps of the methods involved in the above embodiments. For details, please refer to the relevant descriptions in the preceding method embodiments.

[0122] The processor and memory can be separate structures or integrated structures. When the processor and memory are separate structures, they can be coupled together via a bus.

[0123] In this embodiment of the disclosure, a non-transitory computer-readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, implements the steps of a billion-pixel-level target detection method in any of the above embodiments.

[0124] Those skilled in the art will understand that embodiments of this disclosure can be provided as methods, systems, or computer program products. Therefore, this disclosure can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this disclosure can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0125] This disclosure is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0126] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0127] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0128] Although preferred embodiments of this disclosure have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this disclosure.

[0129] Obviously, those skilled in the art can make various modifications and variations to this disclosure without departing from its spirit and scope. Therefore, if such modifications and variations fall within the scope of the claims of this disclosure and their equivalents, this disclosure is also intended to include such modifications and variations.

Claims

1. A method for detecting targets at the billion-pixel level, characterized in that, include: A pre-trained fast detector is used to detect and process billion-pixel level images to determine the target size distribution of the billion-pixel level images; A pre-trained SAM model is used to perform background removal processing on the billion-pixel image to determine the billion-pixel image region without background. The pre-trained SAM model is the Segment Anything Model. Based on the target size distribution of the billion-pixel-level image, the pre-trained fast detector and the pre-trained slow detector are used to perform target detection processing on the background-free billion-pixel-level image region to determine the target detection result. The pre-trained fast detector is used to detect larger target regions, and the pre-trained slow detector is used to detect smaller target regions.

2. The method according to claim 1, characterized in that, The process of using a pre-trained fast detector to detect and process billion-pixel-level images to determine the target size distribution of the billion-pixel-level images includes: The pre-trained fast detector is used to perform a first downsampling process on the billion-pixel-level image to determine the first downsampled image; The pre-trained fast detector is used to perform target detection processing on the billion-pixel image to determine the target size distribution of the billion-pixel image.

3. The method according to claim 1, characterized in that, The process of using a pre-trained SAM model to remove background from the billion-pixel image to determine background-free billion-pixel image regions includes: The pre-trained SAM model is used to perform a second downsampling process on the billion-pixel image to determine the second downsampled image; Based on the preset target threshold, the pre-trained SAM model is used to perform region filtering on the second downsampled image to determine the background-free billion-pixel-level image region.

4. The method according to claim 3, characterized in that, The step of using the pre-trained SAM model to perform region filtering processing on the second downsampled image according to a preset target threshold to determine the background-free billion-pixel-level image region includes: Pixels in the second downsampled image that are larger than the preset target threshold are identified as background regions and removed. Pixels in the second downsampled image that are smaller than the preset target threshold are identified as the background-free billion-pixel-level image region and retained.

5. The method according to claim 1, characterized in that, The step of performing target detection processing on the background-free billion-pixel image region using the pre-trained fast detector and the pre-trained slow detector based on the target size distribution of the billion-pixel-level image, and determining the target detection result, includes: The pre-trained fast detector is optimized using TensorRT to determine the optimized fast detector. The pre-trained slow detector is converted to half-floating-point precision to determine the converted slow detector. Based on the preset detection threshold and the target size distribution of the billion-pixel image, the optimized fast detector and the converted slow detector are used to perform target detection processing on the background-free billion-pixel image region to determine the target detection result.

6. The method according to claim 5, characterized in that, The step of performing target detection processing on the background-free billion-pixel image region using the optimized fast detector and the converted slow detector, based on a preset detection threshold and the target size distribution of the billion-pixel-level image, and determining the target detection result, includes: If the target size of the target in the backgroundless billion-pixel-level image region is larger than the preset detection threshold, the optimized fast detector is used to perform target detection processing to determine the first target detection result. If the target size of the target in the backgroundless billion-pixel-level image region is not greater than the preset detection threshold, the target detection process is performed using the converted slow detector to determine the second target detection result.

7. The method according to claim 1, characterized in that, The method further includes: Preprocess the pre-defined dataset to determine the training dataset; The training dataset is used to train a preset fast detector, a preset slow detector, and a preset SAM model to determine the pre-trained fast detector, the pre-trained slow detector, and the pre-trained SAM model.

8. A billion-pixel-level target detection device, characterized in that, include: The first detection and processing module is used to perform detection processing on the billion-pixel-level image using a pre-trained fast detector to determine the target size distribution of the billion-pixel-level image. The background processing module is used to perform background removal processing on the billion-pixel image using a pre-trained SAM model to determine the billion-pixel image region without background, wherein the pre-trained SAM model is the Segment AnythingModel model. The second detection processing module is used to perform target detection processing on the background-free billion-pixel image region using the pre-trained fast detector and the pre-trained slow detector according to the target size distribution of the billion-pixel image, and determine the target detection result. The pre-trained fast detector is used to detect larger target regions, and the pre-trained slow detector is used to detect smaller target regions.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method described in any one of claims 1-7.

10. An electronic device, characterized in that, include: A memory on which computer programs are stored; A processor for executing the computer program in the memory to implement the steps of the method according to any one of claims 1-7.

Images