WO2023213178A1 - 一种高分辨率遥感图像的目标检测方法及装置 - Google Patents

一种高分辨率遥感图像的目标检测方法及装置 Download PDF

Info

Publication number
WO2023213178A1
WO2023213178A1 PCT/CN2023/088349 CN2023088349W WO2023213178A1 WO 2023213178 A1 WO2023213178 A1 WO 2023213178A1 CN 2023088349 W CN2023088349 W CN 2023088349W WO 2023213178 A1 WO2023213178 A1 WO 2023213178A1
Authority
WO
WIPO (PCT)
Prior art keywords
target
detection
area
remote sensing
inspection
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2023/088349
Other languages
English (en)
French (fr)
Inventor
刘振德
张吴越
刘鹏
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Zhejiang Lab
Original Assignee
Zhejiang Lab
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Zhejiang Lab filed Critical Zhejiang Lab
Priority to US18/464,249 priority Critical patent/US12014536B2/en
Publication of WO2023213178A1 publication Critical patent/WO2023213178A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/762—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using clustering, e.g. of similar faces in social networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/20—Analysis of motion
    • G06T7/246—Analysis of motion using feature-based methods, e.g. the tracking of corners or segments
    • G06T7/248—Analysis of motion using feature-based methods, e.g. the tracking of corners or segments involving reference images or patches
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/60—Analysis of geometric attributes
    • G06T7/62—Analysis of geometric attributes of area, perimeter, diameter or volume
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/20—Image preprocessing
    • G06V10/255—Detecting or recognising potential candidate objects based on visual cues, e.g. shapes
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/20—Image preprocessing
    • G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/62—Extraction of image or video features relating to a temporal dimension, e.g. time-based feature extraction; Pattern tracking
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10016—Video; Image sequence
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10032—Satellite or aerial image; Remote sensing
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20081—Training; Learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20112—Image segmentation details
    • G06T2207/20132—Image cropping
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G06V20/13—Satellite images
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V2201/00—Indexing scheme relating to image or video recognition or understanding
    • G06V2201/07—Target detection

Definitions

  • This application relates to the field of deep learning technology and remote sensing image processing, and in particular to a target detection method and device for high-resolution remote sensing images.
  • Remote sensing satellites are generally located at an altitude of 300-2,000 kilometers and have an orbital period of about 90 minutes.
  • High-resolution remote sensing can accurately observe targets of various sizes non-uniformly distributed on the earth's surface with meter-level or even sub-meter-level spatial resolution.
  • High-resolution remote sensing image data has been used in many fields such as surveying and mapping, urban planning, transportation, water conservancy, agriculture, forestry, environmental resource monitoring, and military.
  • all remote sensing images cannot be transmitted back to the ground for processing.
  • Spaceborne or airborne platforms often process remote sensing image data in real time through the edge computing platform they carry, and then transmit the results or key areas back for processing.
  • the present invention proposes a target detection method that uses different models for different areas after adaptively dividing high-resolution remote sensing images, thereby achieving the purpose of improving the image target detection accuracy of the computing system and improving the energy efficiency of the system.
  • the purpose of the embodiments of this application is to provide a target detection method and device for high-resolution remote sensing images, so as to solve the technical problems of large resource usage and low computational efficiency existing in related technologies.
  • a target detection method for high-resolution remote sensing images is provided, which is applied on a computing platform, including:
  • adaptive image partitioning is performed on the original high-resolution remote sensing image to obtain different cluster inspection areas, and then the required inspection areas and non-inspection areas are obtained;
  • Offline training is suitable for target detection models of different densities and sizes, and the delays and detection accuracy of different models on the computing platform are measured, and the models and corresponding delays and detection accuracy are loaded on the computing platform.
  • the target detection results it also includes:
  • the historical detection results stored on the computing platform are updated.
  • adaptive image partitioning is performed on the original high-resolution remote sensing image to obtain different cluster inspection areas, and then the required inspection areas and non-inspection areas are obtained, including:
  • S21 Cluster according to the target and the target's position, and calculate the initial value of the number of clustering categories, where the target information includes the target's position, size and category;
  • S25 Combined with the motion information of the target, calculate the area where the target within the cluster may appear, and set it as the cluster waiting area. All cluster waiting areas together form the must-inspection area.
  • the original high-resolution remote sensing image is not a must-inspection area.
  • the area is a non-inspection area, where the movement information of the target is stored on the computing platform.
  • model selection is performed for the must-inspection area and a target detection plan for the must-inspection area is generated, including:
  • S32 Evenly divide one of the cluster waiting areas for inspection according to different input sizes of the models in the model set, and generate at least one model combination that is feasible for the cluster waiting area;
  • a target detection device for high-resolution remote sensing images is provided, which is applied on a computing platform and includes:
  • the first acquisition module is used to acquire original high-resolution remote sensing images from the sensor
  • the second acquisition module is used to acquire the target information in the area corresponding to the original high-resolution remote sensing image and the extended area in which the area extends a predetermined distance to the surroundings from the historical detection results;
  • a partitioning module configured to perform adaptive image partitioning on the original high-resolution remote sensing image according to the target information, to obtain different cluster waiting areas for inspection, and then to obtain mandatory inspection areas and non-obligatory inspection areas;
  • the generation module is used to select the model of the must-inspection area and generate the target detection plan for the must-inspection area;
  • the execution module is used to execute the target detection plan and obtain the detection results of each required inspection area
  • the detection module is used to determine whether the computing platform has redundant computing resources to detect non-compulsory inspection areas. If so, perform dynamic partition detection of the non-compulsory inspection areas to obtain the detection results of the non-compulsory inspection areas. The detection results and the detection results of the non-compulsory detection areas are combined into the target detection results, otherwise the detection results of the required detection areas are used as the target detection results.
  • an electronic device including:
  • processors one or more processors
  • Memory used to store one or more programs
  • the one or more processors When the one or more programs are executed by the one or more processors, the one or more processors are caused to implement the method described in the first aspect.
  • a computer-readable storage medium on which computer instructions are stored. When the instructions are executed by a processor, the steps of the method described in the first aspect are implemented.
  • this application refers to historical detection results to obtain target information in the corresponding area and extended area of the image, thereby reducing the computing resources used when obtaining target information; based on the target information, adaptive partitioning and must-detection areas are implemented.
  • priority is given to completing the detection of required inspection areas to avoid wasting resources in detecting untargeted areas, improving the overall detection accuracy and speeding up target detection.
  • Figure 1 is a flow chart of a target detection method for high-resolution remote sensing images according to an exemplary embodiment.
  • FIG. 2 is a flowchart of step S13 according to an exemplary embodiment.
  • FIG. 3 is a flowchart of step S14 according to an exemplary embodiment.
  • Figure 4 is a flow chart of a target detection method for high-resolution remote sensing images according to an exemplary embodiment.
  • Figure 5 is a block diagram of a target detection device for high-resolution remote sensing images according to an exemplary embodiment.
  • FIG. 6 is a schematic diagram of an electronic device according to an exemplary embodiment.
  • first, second, third, etc. may be used in this application to describe various information, the information should not be limited to these terms. These terms are only used to distinguish information of the same type from each other.
  • first information may also be called second information, and similarly, the second information may also be called first information.
  • word “if” as used herein may be interpreted as "when” or “when” or “in response to determining.”
  • Figure 1 is a flow chart of a target detection method for high-resolution remote sensing images according to an exemplary embodiment. As shown in Figure 1, the method is applied to a computing platform and may include the following steps:
  • Step S11 Obtain original high-resolution remote sensing images from the sensor
  • Step S12 Obtain the target information in the area corresponding to the original high-resolution remote sensing image and the extended area in which the area extends a predetermined distance to the surroundings from the historical detection results;
  • Step S13 According to the target information, perform adaptive image partitioning on the original high-resolution remote sensing image to obtain different cluster inspection areas, and then obtain the required inspection area and the non-inspection area;
  • Step S14 Carry out model selection for the must-inspection area and generate a target detection plan for the must-inspection area;
  • Step S15 Execute the target detection plan to obtain the detection results of the required inspection area
  • Step S16 Determine whether the computing platform has redundant computing resources to detect the non-compulsory inspection areas. If so, perform dynamic partition detection of the non-compulsory inspection areas, obtain the detection results of the non-compulsory inspection areas, and combine the detection results of the required inspection areas. The detection results of the non-compulsory detection area are combined into the target detection results, otherwise the detection results of the mandatory detection area are used as the target detection results.
  • this application refers to historical detection results to obtain target information in the corresponding area and extended area of the image, thereby reducing the computing resources used when obtaining target information; based on the target information, adaptive partitioning and must-detection areas are implemented.
  • priority is given to completing the detection of required inspection areas to avoid wasting resources in detecting untargeted areas, improving the overall detection accuracy and speeding up target detection.
  • the computing platform refers to hardware devices that can run deep neural network models, including but not limited to artificial intelligence processors, CPUs, GPUs, FPGAs and other devices.
  • step S11 it may also include:
  • Offline training is suitable for target detection models of different densities and sizes, and the delays and detection accuracy of different models on the computing platform are measured, and the models and corresponding delays and detection accuracy are loaded on the computing platform.
  • models with different target densities and different input sizes are trained on an offline computer for remote sensing target detection, the trained model is loaded into the computing platform, and images of corresponding sizes and densities are input, and the experimental test results are Latency and detection accuracy of each model on this computing platform.
  • Offline training and experimental testing of each model's latency and accuracy are conducive to accurately budgeting the computing resources required by the model during actual operation, so that detection results with the best accuracy can be obtained under limited computing resources.
  • the computing platform is also provided with target motion information, side length thresholds and density thresholds for detecting whether the clustering results are reasonable. Based on the actual target motion information, the area of the must-detect area, side length thresholds and density can be more accurately determined.
  • the setting of thresholds can determine whether the results of target clustering are reasonable and avoid clustering being affected by outliers and local high densities. Their settings can effectively avoid missed detections and waste of computing resources for over-detection.
  • step S11 the original high-resolution remote sensing image is obtained from the sensor
  • the high-resolution images captured by the optical and radar imaging loads carried by the satellite are read into the storage space of the computing platform.
  • step S12 target information in the area corresponding to the original high-resolution remote sensing image and the extended area in which the area extends a predetermined distance to the surroundings is obtained from the historical detection results;
  • target detection information of historical remote sensing images corresponding to the same geographical location is obtained, including target size, geographical location, and target category.
  • target detection information of the remote sensing image of the extended area extending a predetermined distance from the historical area to the surroundings is obtained, including the target size, geographical location, and target category.
  • the method for extending the area by a predetermined distance to all sides is to multiply the set target movement speed threshold by the movement cycle duration of the satellite corresponding to the original high-resolution remote sensing image to obtain the movement distance of the target within a cycle, so
  • the expansion area is an area in which the area extends to all sides by a periodic target movement distance.
  • the entire image is regarded as a non-must-detection area, and dynamic partition detection of the non-must-detection area is performed.
  • step S13 according to the target information, adaptive image partitioning is performed on the original high-resolution remote sensing image to obtain different cluster waiting areas, and then the required inspection areas and non-inspection areas are obtained;
  • this step can include the following sub-steps:
  • S21 Cluster according to the target and the target's position, and calculate the initial value of the number of clustering categories, where the target information includes the target's position, size and category;
  • adaptive partitioning of remote sensing images is achieved through a clustering algorithm, in which the initial value of the number of clustering categories can be calculated based on the number of targets in the area in historical information.
  • the clustering algorithm can include k-means clustering, hierarchical Clustering and other algorithms.
  • a clustering algorithm is executed, and the category result obtained by each cluster is regarded as a cluster.
  • This step can cluster all possible targets in the remote sensing image and assist in densely identifying the partitions to which the target belongs. sexual judgment.
  • a circumscribed rectangle is constructed for each cluster obtained after clustering.
  • the horizontal distance between the two furthest targets in the horizontal direction within the cluster is the width of the circumscribed rectangle.
  • the vertical distance between the two furthest targets in the vertical direction within the cluster is The height of the circumscribed rectangle.
  • the constructed circumscribed rectangle should be the smallest rectangle containing all points in the cluster.
  • the target density of the cluster is calculated through the side length of the circumscribed rectangle and the number of targets within the rectangle.
  • the preset density threshold by comparing the preset density threshold with the density of the target within the circumscribed rectangle, if the density within the rectangle has exceeded the density threshold, but one of the side lengths of the circumscribed rectangle exceeds the preset side length threshold, it indicates that the cluster is
  • the targets are too dispersed in the horizontal or vertical direction and need to be re-clustered by updating the number of clustered categories. Judging the distribution of circumscribed rectangular targets can verify the effectiveness of clustering and prevent misclassification of some outliers.
  • S25 Combined with the motion information of the target, calculate the area where the target within the cluster may appear, and set it as the cluster waiting area. All cluster waiting areas together form the must-inspection area.
  • the original high-resolution remote sensing image is not a must-inspection area.
  • the area is a non-inspection area, where the movement information of the target is stored on the computing platform;
  • the position of each target in the cluster can be motion expanded to obtain the expanded target area, which is the cluster. inspection area. After all clusters are expanded, the inspection waiting areas of all clusters together form the required inspection area.
  • the area in the original high-resolution remote sensing image other than the required inspection area is a non-inspection area.
  • the way for the cluster to expand based on the movement of the targets in the cluster is to obtain the movement of each target in the cluster within a cycle based on the movement speed of each target in the cluster multiplied by the movement period duration of the satellite corresponding to the high-resolution remote sensing image. distance.
  • the expanded area is an area that is expanded by a periodic movement distance of each target in the cluster based on the geographical location information of each target in the cluster. The expansion of the area where the target may exist can prevent missed detection due to target movement leaving the original area.
  • the excess area will be recorded and used as a required inspection area for detection when calculating adjacent images.
  • step S14 model selection for the must-inspection area is performed, and a target detection plan for the must-inspection area is generated;
  • this step may include the following sub-steps:
  • the information of all models in the computing platform is traversed, and a model set that can be used to calculate the density and model size is retrieved.
  • a model set that can be used to calculate the density and model size is retrieved.
  • S32 Evenly divide one of the cluster waiting areas for inspection according to different input sizes of the models in the model set, and generate at least one model combination that is feasible for the cluster waiting area;
  • the size of the cluster detection area is 1024*1024
  • a model with an input size of 1024*1024 can be used, or four models of 512*512 can be used.
  • each cluster waiting area has one or more available models. After combining the available models of all cluster waiting areas, one or more target detection model solutions are constructed for the entire image.
  • the total delay is weighted and summed up with the number of uses as the weight of the latencies of all models, and the total accuracy is weighted with the number of uses as the weight of all models.
  • Accuracy is weighted average
  • model solutions are traversed in the order of total accuracy from high to low, and the model solution in which the total delay is less than the computing platform delay limit is selected. If there are solutions with the same total accuracy, the model solution with the smallest total delay is preferred.
  • the selected model scheme ensures the highest detection accuracy without exceeding the latency limit of the computing platform.
  • step S15 the target detection plan is executed to obtain the detection results of the required inspection area
  • the model scheme has been determined, which also determines the models to be used in different areas of the image. If the computing platform supports parallel computing, each model and the corresponding image area can be executed in parallel. If parallel computing is not supported, each model and corresponding image area is executed serially.
  • step S16 it is determined whether the computing platform has redundant computing resources to detect the non-compulsory inspection areas. If so, dynamic partition detection of the non-compulsory inspection areas is performed to obtain the detection results of the non-compulsory inspection areas. The detection results of the detection area and the detection results of the non-obligatory detection area are combined into the target detection results, otherwise the detection results of the required detection area are used as the target detection results.
  • the delay remainder is calculated based on the constraints of the computing platform minus the total delay of the executed model scenario.
  • the non-inspection area will not be detected.
  • the non-inspection area will be detected.
  • the dynamic programming algorithm is used to calculate the model solution that maximizes the detection area under the constraints of the boundary and delay residual of the non-compulsory detection area.
  • the model solution can be executed in parallel. If parallel computing is not supported, each model in the model scheme is executed serially. And merge the test results of non-compulsory inspection areas with the test results of mandatory inspection areas.
  • This design is to maximize the area of non-necessary detection areas under the delay residual limit, which can reduce the possibility of missed detection of targets and maximize the use of computing resources.
  • step S16 after obtaining the target detection result through step S16, it may also include:
  • Step S17 Update the historical detection results stored on the computing platform according to the target detection results.
  • the targets in the corresponding area of the image are detected through the above calculations, all previous targets corresponding to the area are deleted from the historical detection results stored on the computing platform, and the latest target detection results are written.
  • the updated historical detection results can also be transmitted to the satellite computer system for processing.
  • this application also provides embodiments of a target detection device for high-resolution remote sensing images.
  • FIG. 5 is a block diagram of a target detection device for high-resolution remote sensing images according to an exemplary embodiment.
  • the device includes:
  • the first acquisition module 21 is used to acquire original high-resolution remote sensing images from the sensor;
  • the second acquisition module 22 is used to acquire the target information in the area corresponding to the original high-resolution remote sensing image and the extended area in which the area extends a predetermined distance to the surroundings from the historical detection results;
  • the partitioning module 23 is used to perform adaptive image partitioning on the original high-resolution remote sensing image according to the target information to obtain different cluster waiting areas for inspection, and then obtain the required inspection area and the non-obligatory inspection area;
  • the generation module 24 is used to select a model for the must-inspection area and generate a target detection plan for the must-inspection area;
  • the execution module 25 is used to execute the target detection plan and obtain the detection results of each required inspection area
  • the detection module 26 is used to determine whether the computing platform has redundant computing resources to detect non-compulsory inspection areas. If so, perform dynamic partition detection of the non-compulsory inspection areas, obtain the detection results of the non-compulsory inspection areas, and combine the required inspection areas with each other. The detection results of the required detection areas and the detection results of the non-obligatory detection areas are combined into the target detection results, otherwise the detection results of each of the mandatory detection areas are used as the target detection results.
  • the device embodiment since it basically corresponds to the method embodiment, please refer to the partial description of the method embodiment for relevant details.
  • the device embodiments described above are only illustrative.
  • the units described as separate components may or may not be physically separated.
  • the components shown as units may or may not be physical units, that is, they may be located in One location, or it can be distributed across multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this application. Persons of ordinary skill in the art can understand and implement the method without any creative effort.
  • this application also provides an electronic device, including: one or more processors; a memory for storing one or more programs; when the one or more programs are executed by the one or more processors , so that the one or more processors implement the target detection method of high-resolution remote sensing images as mentioned above.
  • Figure 6 it is a hardware structure diagram of any device with data processing capabilities where a target detection method for high-resolution remote sensing images provided by an embodiment of the present invention is located.
  • any device with data processing capabilities where the device in the embodiment is located may also include other hardware based on the actual functions of the device with data processing capabilities, which will not be described again.
  • this application also provides a computer-readable storage medium on which computer instructions are stored.
  • the computer-readable storage medium may be an internal storage unit of any device with data processing capabilities as described in any of the foregoing embodiments, such as a hard disk or a memory.
  • the computer-readable storage medium may also be an external storage device of the wind turbine, such as a plug-in hard disk, a smart media card (SMC), an SD card, or a flash card equipped on the device. wait.
  • the computer-readable storage medium may also include both an internal storage unit and an external storage device of any device with data processing capabilities.
  • the computer-readable storage medium is used to store the computer program and other programs and data required by any device with data processing capabilities, and can also be used to temporarily store data that has been output or is to be output.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Multimedia (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Artificial Intelligence (AREA)
  • Medical Informatics (AREA)
  • Databases & Information Systems (AREA)
  • Geometry (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Molecular Biology (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Image Analysis (AREA)

Abstract

本发明公开了一种高分辨率遥感图像的目标检测方法及装置，该方法包括：从传感器获取原始高分辨率遥感图像；从历史检测结果中获取原始高分辨率遥感图像所对应区域以及所述区域向四周扩展预定距离的扩展区域内的目标信息；根据目标信息，对原始高分辨率遥感图像进行自适应的图像分区，得到不同的簇侯检区，进而得到必检区和非必检区；进行必检区的模型选择，生成必检区的目标检测方案；执行目标检测方案，得到必检区的检测结果；判断计算平台是否有多余计算资源以检测非必检区，若有，则进行非必检区的动态分区检测，得到非必检区的检测结果，将必检区的检测结果和非必检区的检测结果合并为目标检测结果，否则将必检区的检测结果作为目标检测结果。

Description

一种高分辨率遥感图像的目标检测方法及装置 技术领域
本申请涉及深度学习技术、遥感图像处理领域，尤其涉及一种高分辨率遥感图像的目标检测方法及装置。
 背景技术
随着传感器技术和航天遥感技术的进步，针对遥感图像目标检测的需求不断提升。遥感卫星一般所处高度在300-2000千米，轨道周期在90分钟左右。高分辨率遥感可以以米级甚至亚米级空间分辨率精细观测非均匀分布在地球表面的各种尺寸目标。高分辨率遥感图像数据在测绘制图、城市规划、交通、水利、农业、林业、环境资源监测、军事等领域得到了众多应用。但是由于高分辨率图像较大和空地通信的带宽限制，不能将所有遥感图像回传至地面处理。星载或机载平台往往通过搭载的边缘计算平台先对遥感图像数据做实时处理，再将结果或者重点区域回传处理。
使用深度学习目标检测模型处理图像的技术已经十分常见。大多数目标检测模型都设计了低分辨率的输入，高分辨率图像目标检测的常见做法是调整图像大小，进行图像分割或扩大网络的输入分辨率。调整图像大小和图像分割会影响精度。
在实现本发明的过程中，发明人发现现有技术中至少存在如下问题：
扩大输入分辨率会导致更慢的模型训练和推理的速度，也会增加单位时间功耗。这些处理方法在资源有限的的边缘计算平台是不适用的。本发明提出了一种对高分辨率遥感图像进行自适应划分后，针对不同的区域使用不同模型的目标检测方法，从而达到提高计算系统的图像目标检测精度和提升系统能效的目的。
 发明内容
本申请实施例的目的是提供一种高分辨率遥感图像的目标检测方法及装置，以解决相关技术中存在的占用资源大、计算效率低的技术问题。
根据本申请实施例的第一方面，提供一种高分辨率遥感图像的目标检测方法，应用于计算平台上，包括：
从传感器获取原始高分辨率遥感图像；
从历史检测结果中获取所述原始高分辨率遥感图像所对应区域以及所述区域向四周扩展预定距离的扩展区域内的目标信息；
根据所述目标信息，对所述原始高分辨率遥感图像进行自适应的图像分区，得到不同的簇侯检区，进而得到必检区和非必检区；
进行必检区的模型选择，生成必检区的目标检测方案；
执行所述目标检测方案，得到必检区的检测结果；
判断计算平台是否有多余计算资源以检测非必检区，若有，则进行非必检区的动态分区检测，得到非必检区的检测结果，将所述必检区的检测结果和所述非必检区的检测结果合并为目标检测结果，否则将所述必检区的检测结果作为目标检测结果。
进一步地，从传感器获取原始高分辨率遥感图像之前还包括：
离线训练适合不同密度和尺寸的目标检测模型，并测量出不同模型在计算平台上的延迟和检测精度，并将模型及对应的延迟和检测精度装载在所述计算平台上。
进一步地，在得到目标检测结果之后还包括：
根据所述目标检测结果，更新计算平台上存储的历史检测结果。
进一步地，根据所述目标信息，对所述原始高分辨率遥感图像进行自适应的图像分区，得到不同的簇侯检区，进而得到必检区和非必检区，包括：
S21：根据目标和目标的位置进行聚类，计算聚类的类别数的初始值，其中所述目标信息包括目标的位置、尺寸和类别；
S22：根据所述类别数的初始值，执行聚类算法，得到对应数量的簇；
S23：对每个簇构建一个外接矩形，计算该矩形的边长以及矩形内目标的密度；
S24：如果矩形的边长超过边长阈值且矩形内目标的密度超过密度阈值，更新所述类别数的值，返回步骤S22重新聚类；反之则根据目标的密度和尺寸，选择能够处理所述密度和尺寸的模型集合；
S25：结合目标的运动信息，计算出簇内目标可能出现的区域，设为簇侯检区，所有簇侯检区共同组成必检区，所述原始高分辨率遥感图像中不为必检区的区域则为非必检区，其中所述目标的运动信息存储在所述计算平台上。
进一步地，进行必检区的模型选择，生成必检区的目标检测方案，包括：
S31：根据目标的密度和尺寸，选择能够处理所述密度和尺寸的模型集合；
S32：根据所述模型集合中模型的不同输入尺寸将其中一个簇侯检区均匀分割，产生对该簇侯检区可行的至少一种模型组合；
S33：对所述必检区中的每一个簇侯检区，执行S32，将所有簇侯检区的模型组合之后形成至少一种模型方案；
S34：根据每个模型方案的精度和延迟信息，计算每个模型方案对于检测所述必检区的精度和总延迟；
S35：在所有模型方案中以整体精度优先方式选择总延迟满足所述计算平台的延迟限制的方案作为必检区的目标检测方案。
根据本申请实施例的第二方面，提供一种高分辨率遥感图像的目标检测装置，应用于计算平台上，包括：
第一获取模块，用于从传感器获取原始高分辨率遥感图像；
第二获取模块，用于从历史检测结果中获取所述原始高分辨率遥感图像所对应区域以及所述区域向四周扩展预定距离的扩展区域内的目标信息；
分区模块，用于根据所述目标信息，对所述原始高分辨率遥感图像进行自适应的图像分区，得到不同的簇侯检区，进而得到必检区和非必检区；
生成模块，用于进行必检区的模型选择，生成必检区的目标检测方案；
执行模块，用于执行所述目标检测方案，得到各必检区的检测结果；
检测模块，用于判断计算平台是否有多余计算资源以检测非必检区，若有，则进行非必检区的动态分区检测，得到非必检区的检测结果，将所述各必检区的检测结果和所述非必检区的检测结果合并为目标检测结果，否则将所述各必检区的检测结果作为目标检测结果。
根据本申请实施例的第三方面，提供一种电子设备，包括：
一个或多个处理器；
存储器，用于存储一个或多个程序；
当所述一个或多个程序被所述一个或多个处理器执行，使得所述一个或多个处理器实现如第一方面所述的方法。
根据本申请实施例的第四方面，提供一种计算机可读存储介质，其上存储有计算机指令，该指令被处理器执行时实现如第一方面所述方法的步骤。
本申请的实施例提供的技术方案可以包括以下有益效果：
由上述实施例可知，本申请参考历史检测结果，获取图像对应区域及扩展区域内的目标信息，从而减少了获取目标信息时使用的计算资源；根据目标信息，实现自适应分区和必检区、非必检区的分级处理，优先完成必检区检测，避免检测无目标区域浪费资源，提升整体检测精度和加快目标检测的速度。
应当理解的是，以上的一般描述和后文的细节描述仅是示例性和解释性的，并不能限制本申请。
 附图说明
此处的附图被并入说明书中并构成本说明书的一部分，示出了符合本申请的实施例，并与说明书一起用于解释本申请的原理。
图1是根据一示例性实施例示出的一种高分辨率遥感图像的目标检测方法的流程图。
图2是根据一示例性实施例示出的步骤S13的流程图。
图3是根据一示例性实施例示出的步骤S14的流程图。
图4是根据一示例性实施例示出的一种高分辨率遥感图像的目标检测方法的流程图。
图5是根据一示例性实施例示出的一种高分辨率遥感图像的目标检测装置的框图。
图6是根据一示例性实施例示出的一种电子设备的示意图。
 具体实施方式
这里将详细地对示例性实施例进行说明，其示例表示在附图中。下面的描述涉及附图时，除非另有表示，不同附图中的相同数字表示相同或相似的要素。以下示例性实施例中所描述的实施方式并不代表与本申请相一致的所有实施方式。相反，它们仅是与如所附权利要求书中所详述的、本申请的一些方面相一致的装置和方法的例子。
在本申请使用的术语是仅仅出于描述特定实施例的目的，而非旨在限制本申请。在本申请和所附权利要求书中所使用的单数形式的“一种”、“所述”和“该”也旨在包括多数形式，除非上下文清楚地表示其他含义。还应当理解，本文中使用的术语“和/或”是指并包含一个或多个相关联的列出项目的任何或所有可能组合。
应当理解，尽管在本申请可能采用术语第一、第二、第三等来描述各种信息，但这些信息不应限于这些术语。这些术语仅用来将同一类型的信息彼此区分开。例如，在不脱离本申请范围的情况下，第一信息也可以被称为第二信息，类似地，第二信息也可以被称为第一信息。取决于语境，如在此所使用的词语“如果”可以被解释成为“在……时”或“当……时”或“响应于确定”。
图1是根据一示例性实施例示出的一种高分辨率遥感图像的目标检测方法的流程图，如图1所示，该方法应用于计算平台中，可以包括以下步骤：
步骤S11：从传感器获取原始高分辨率遥感图像；
步骤S12：从历史检测结果中获得所述原始高分辨率遥感图像所对应区域以及所述区域向四周扩展预定距离的扩展区域内的目标信息；
步骤S13：根据所述目标信息，对所述原始高分辨率遥感图像进行自适应的图像分区，得到不同的簇侯检区，进而得到必检区和非必检区；
步骤S14：进行必检区的模型选择，生成必检区的目标检测方案；
步骤S15：执行所述目标检测方案，得到必检区的检测结果；
步骤S16：判断计算平台是否有多余计算资源以检测非必检区，若有，则进行非必检区的动态分区检测，得到非必检区的检测结果，将所述必检区的检测结果和所述非必检区的检测结果合并为目标检测结果，否则将所述必检区的检测结果作为目标检测结果。
由上述实施例可知，本申请参考历史检测结果，获取图像对应区域及扩展区域内的目标信息，从而减少了获取目标信息时使用的计算资源；根据目标信息，实现自适应分区和必检区、非必检区的分级处理，优先完成必检区检测，避免检测无目标区域浪费资源，提升整体检测精度和加快目标检测的速度。
在具体实施中，计算平台指的是可运行深度神经网络模型的硬件设备，包含且不限于人工智能处理器，CPU，GPU，FPGA等设备。
在步骤S11之前，还可以包括：
离线训练适合不同密度和尺寸的目标检测模型，并测量出不同模型在计算平台上的延迟和检测精度，并将模型及对应的延迟和检测精度装载在所述计算平台上。
具体地， 在离线计算机上对不同目标密度和不同输入尺寸的模型正对遥感目标检测做训练，将训练好的模型在装载在计算平台内，并在输入对应尺寸和密度的图像，实验测试出每个模型在该计算平台的延迟和检测精度。离线训练和实验测试每个模型延迟和精度有利于在实际运行时准确预算模型所需的计算资源，使得在有限的计算资源下获得精度最佳的检测结果。所述计算平台上还设置有目标运动信息、用于检测聚类结果是否合理的边长阈值和密度阈值，根据实际的目标运动信息可以更加准确判断出必检区的区域，边长阈值和密度阈值的设置可以判断目标聚类的结果是否合理，避免聚类时受异常点和局部高密度的影响，他们的设置可以有效避免漏检和浪费计算资源过度检测。
在步骤S11的具体实施中，从传感器获取原始高分辨率遥感图像；
具体地，通过卫星搭载的光学、雷达成像载荷拍摄到的高分辨率图像，读取到计算平台的存储空间内。
在步骤S12的具体实施中，从历史检测结果中获得所述原始高分辨率遥感图像所对应区域以及所述区域向四周扩展预定距离的扩展区域内的目标信息；
具体地，根据所述原始高分辨率图像的地理位置，获取对应相同地理位置的历史遥感图像的目标检测信息，包括目标的大小、地理位置、目标类别。同时根据所述原始高分辨率图像的地理位置，获取历史所述区域向四周扩展预定距离的扩展区域的遥感图像的目标检测信息，包括目标的大小、地理位置、目标类别。其中，所述区域向四周扩展预定距离的方式为：将设定的目标运动速度阈值乘以所述原始高分辨率遥感图像对应的卫星的运动周期时长，获得一个周期内目标的运动距离，所述扩展区域即为所述区域向四周扩展一个周期目标运动距离的区域。通过该设计，可以结合历史检测信息对该周期遥感图像内的可能存在目标位置做预测，辅助自适应分区的算法，同时结合扩展区域内的目标检测信息，可以避免所述区域外向所述区域内运动的新目标信息的遗漏，辅助自适应分区的算法。
在具体实施中，若不存在历史检测结果，则将整张图像视为非必检区，执行非必检区的动态分区检测。
在步骤S13的具体实施中，根据所述目标信息，对所述原始高分辨率遥感图像进行自适应的图像分区，得到不同的簇侯检区，进而得到必检区和非必检区；
具体地，如图2所示，此步骤可以包括以下子步骤：
S21：根据目标和目标的位置进行聚类，计算聚类的类别数的初始值，其中所述目标信息包括目标的位置、尺寸和类别；
具体地，通过聚类算法实现对遥感图像的自适应分区，其中聚类类别数的初始值可根据历史信息中该区域的目标数量计算得到，其中聚类算法可以包括k-means聚类、层次聚类等算法。
S22：根据所述类别数的初始值，执行聚类算法，得到对应数量的簇；
具体地，根据所述类别数的初始值，执行聚类算法，每一个聚类得到的类别结果作为一个簇，该步骤可以对遥感图像内所有的可能目标进行聚类，辅助对目标所属分区密集性的判断。
S23：对每个簇构建一个外接矩形，计算该矩形的边长以及矩形内目标的密度；
具体地，对每个聚类后得到的簇构建一个外接矩形，在簇内水平方向最远两个目标的水平距离为外接矩形的宽，以簇内垂直向最远两个目标的垂直距离为外接矩形的高，构建的外接矩形应为包含簇内所有点的最小矩形，通过该外接矩形的边长以及矩形内目标的数量计算该簇的目标密度。
S24：如果矩形内目标的密度超过密度阈值且矩形的边长超过边长阈值，更新所述类别数的值，返回步骤S202重新聚类；
具体地，通过预设置的密度阈值与外接矩形内目标的密度比较，若矩形内的密度已经超过密度阈值，但外接矩形的其中一条边长超过了预设置的边长阈值，则表明该簇的目标在水平或垂直方向上太过分散，需要通过更新聚类的类别数来重新进行聚类。对外接矩形目标分布的判断，可以验证聚类的有效性，防止一些异常点的误分类。
S25：结合目标的运动信息，计算出簇内目标可能出现的区域，设为簇侯检区，所有簇侯检区共同组成必检区，所述原始高分辨率遥感图像中不为必检区的区域则为非必检区，其中所述目标的运动信息存储在所述计算平台上；
具体地，根据历史检测信息中所述区域内目标的位置信息，以及目标的运动信息，可以对聚类后簇内各目标的位置进行运动扩展，得到扩展后的目标区域，即为该簇侯检区。对所有簇进行扩展后，所有簇的侯检区共同组成必检区。所述原始高分辨率遥感图像中除必检区之外的区域则为非必检区。其中所述簇根据蔟内目标进行运动扩展的方式为：根据簇内各目标的运动速度乘以所述高分辨率遥感图像对应的卫星的运动周期时长，获得一个周期内簇内各目标的运动距离。所述扩展区域即为结合蔟内各目标的地理位置信息，对蔟内各目标进行一个周期运动距离扩展后的区域。对目标可能存在区域的扩展能防止因目标运动离开原所在区域的漏检。
在具体实施中，如果存在簇侯检区超出了当前图像的区域，这种情况下，将对超出区域做记录，作为相邻图像做计算时的必检区进行检测。
在步骤S14的具体实施中，进行必检区的模型选择，生成必检区的目标检测方案；
具体地，如图3所示，此步骤可以包括以下子步骤：
S31：根据矩形内目标的密度和尺寸，选择能够处理所述密度和尺寸的模型集合；
具体地，根据S23中计算出的该矩形的密度和历史信息中获取的该处目标的大小，遍历计算平台内的所有模型的信息，检索出可以用来计算该密度和模型尺寸的模型集合。对不同密度和大小的目标采用对应的模型可以有效提升检测精度和节省计算资源。
S32：根据所述模型集合中模型的不同输入尺寸将其中一个簇侯检区均匀分割，产生对该簇侯检区可行的至少一种模型组合；
具体地，如，簇侯检区大小为1024*1024，可以使用一个输入大小为1024*1024的模型，也可以使用4个512*512的模型。
S33：对所述必检区中的每一个簇侯检区，执行S32，将所有簇侯检区的模型组合之后形成至少一种模型方案；
具体地，每个簇侯检区都有一个或多个可用的模型，将所有的簇侯检区的可用模型组合之后，对整张图像构建出一种或多种目标检测模型方案。
S34：根据每个模型方案的精度和延迟信息，计算每个模型方案对于检测所述必检区的精度和总延迟；
具体地，根据每个模型的延迟和精度，以及每个模型的使用次数，总的延迟以使用次数为权值对所有模型的延迟加权求和，总精度以使用次数为权值对所有模型的精度做加权求平均数
S35：在所有模型方案中以整体精度优先方式选择总延迟满足所述计算平台的延迟限制的方案作为必检区的目标检测方案；
具体地，以总精度从高到低的顺序遍历模型方案，选择模型方案中总延迟小于计算平台延迟限制的模型方案。如果有总精度相同的方案，优先选择总延迟小的模型方案。所选模型方案确保检测精度达到最高，并且不超出计算平台的延迟限制。
在步骤S15的具体实施中，执行所述目标检测方案，得到必检区的检测结果；
具体地，模型方案已确定，也就确定了图像的不同区域所要使用的模型，如果计算平台支持并行计算，可以并行执行各个模型及对应的图像区域。如果不支持并行计算，串行执行各个模型及对应的图像区域。
在步骤S16的具体实施中，判断计算平台是否有多余计算资源以检测非必检区，若有，则进行非必检区的动态分区检测，得到非必检区的检测结果，将所述必检区的检测结果和所述非必检区的检测结果合并为目标检测结果，否则将所述必检区的检测结果作为目标检测结果。
具体地，根据计算平台的限制减去所执行的模型方案的总延迟，计算出延迟剩余。
如果延迟剩余小于计算平台所装载模型中的最小延迟，则不对非必检区进行检测。
如果延迟剩余大于计算平台所装载模型中的最小延迟，则对非必检区进行检测。通过非必检区的边界，延迟剩余，各模型输入尺寸，各模型的延迟作为输入，通过动态规划算法，计算出在非必检区的边界和延迟剩余限制下，最大化检测面积的模型方案，如果计算平台支持并行计算，可以并行执行该模型方案。如果不支持并行计算，串行执行该模型方案中的各个模型。并将非必检区的检测结果和必检区的检测结果合并。
这样设计是为了在延迟剩余限制下最大化检测非必检区的面积，可以降低目标被漏检的可能，并且最大化的利用计算资源。
具体地，如图4所示，在通过步骤S16得到目标检测结果之后，还可以包括：
步骤S17：根据所述目标检测结果，更新计算平台上存储的历史检测结果。
具体地，通过以上计算检测得该图像对应区域内的目标，在计算平台上存储的历史检测结果中将之前该区域对应的目标全部删除，并将最新的目标检测结果写入。在具体实施中还可以将更新后的历史检测结果传输给卫星计算机系统进行处理。
与前述的高分辨率遥感图像的目标检测方法的实施例相对应，本申请还提供了高分辨率遥感图像的目标检测装置的实施例。
图5是根据一示例性实施例示出的一种高分辨率遥感图像的目标检测装置框图。参照图5，该装置包括：
第一获取模块21，用于从传感器获取原始高分辨率遥感图像；
第二获取模块22，用于从历史检测结果中获取所述原始高分辨率遥感图像所对应区域以及所述区域向四周扩展预定距离的扩展区域内的目标信息；
分区模块23，用于根据所述目标信息，对所述原始高分辨率遥感图像进行自适应的图像分区，得到不同的簇侯检区，进而得到必检区和非必检区；
生成模块24，用于进行必检区的模型选择，生成必检区的目标检测方案；
执行模块25，用于执行所述目标检测方案，得到各必检区的检测结果；
检测模块26，用于判断计算平台是否有多余计算资源以检测非必检区，若有，则进行非必检区的动态分区检测，得到非必检区的检测结果，将所述各必检区的检测结果和所述非必检区的检测结果合并为目标检测结果，否则将所述各必检区的检测结果作为目标检测结果。
关于上述实施例中的装置，其中各个模块执行操作的具体方式已经在有关该方法的实施例中进行了详细描述，此处将不做详细阐述说明。
对于装置实施例而言，由于其基本对应于方法实施例，所以相关之处参见方法实施例的部分说明即可。以上所描述的装置实施例仅仅是示意性的，其中所述作为分离部件说明的单元可以是或者也可以不是物理上分开的，作为单元显示的部件可以是或者也可以不是物理单元，即可以位于一个地方，或者也可以分布到多个网络单元上。可以根据实际的需要选择其中的部分或者全部模块来实现本申请方案的目的。本领域普通技术人员在不付出创造性劳动的情况下，即可以理解并实施。
相应的，本申请还提供一种电子设备，包括：一个或多个处理器；存储器，用于存储一个或多个程序；当所述一个或多个程序被所述一个或多个处理器执行，使得所述一个或多个处理器实现如上述的高分辨率遥感图像的目标检测方法。如图6所示，为本发明实施例提供的一种高分辨率遥感图像的目标检测方法所在任意具备数据处理能力的设备的一种硬件结构图，除了图6所示的处理器、内存以及网络接口之外，实施例中装置所在的任意具备数据处理能力的设备通常根据该任意具备数据处理能力的设备的实际功能，还可以包括其他硬件，对此不再赘述。
相应的，本申请还提供一种计算机可读存储介质，其上存储有计算机指令，该指令被处理器执行时实现如上述的高分辨率遥感图像的目标检测方法。所述计算机可读存储介质可以是前述任一实施例所述的任意具备数据处理能力的设备的内部存储单元，例如硬盘或内存。所述计算机可读存储介质也可以是风力发电机的外部存储设备，例如所述设备上配备的插接式硬盘、智能存储卡（Smart Media Card，SMC）、SD卡、闪存卡（Flash Card）等。进一步的，所述计算机可读存储介质还可以既包括任意具备数据处理能力的设备的内部存储单元也包括外部存储设备。所述计算机可读存储介质用于存储所述计算机程序以及所述任意具备数据处理能力的设备所需的其他程序和数据，还可以用于暂时地存储已经输出或者将要输出的数据。
本领域技术人员在考虑说明书及实践这里公开的内容后，将容易想到本申请的其它实施方案。本申请旨在涵盖本申请的任何变型、用途或者适应性变化，这些变型、用途或者适应性变化遵循本申请的一般性原理并包括本申请未公开的本技术领域中的公知常识或惯用技术手段。
应当理解的是，本申请并不局限于上面已经描述并在附图中示出的精确结构，并且可以在不脱离其范围进行各种修改和改变。

Claims (8)

  1. 一种高分辨率遥感图像的目标检测方法，其特征在于，应用于计算平台上，包括：
    从传感器获取原始高分辨率遥感图像；
    从历史检测结果中获取所述原始高分辨率遥感图像所对应区域以及所述区域向四周扩展预定距离的扩展区域内的目标信息；
    根据所述目标信息，对所述原始高分辨率遥感图像进行自适应的图像分区，得到不同的簇侯检区，进而得到必检区和非必检区；
    进行必检区的模型选择，生成必检区的目标检测方案；
    执行所述目标检测方案，得到必检区的检测结果；
    判断计算平台是否有多余计算资源以检测非必检区，若有，则进行非必检区的动态分区检测，得到非必检区的检测结果，将所述必检区的检测结果和所述非必检区的检测结果合并为目标检测结果，否则将所述必检区的检测结果作为目标检测结果。
  2. 根据权利要求1所述的方法，其特征在于，从传感器获取原始高分辨率遥感图像之前还包括：
    离线训练适合不同密度和尺寸的目标检测模型，并测量出不同模型在计算平台上的延迟和检测精度，并将模型及对应的延迟和检测精度装载在所述计算平台上。
  3. 根据权利要求1所述的方法，其特征在于，在得到目标检测结果之后还包括：
    根据所述目标检测结果，更新计算平台上存储的历史检测结果。
  4. 根据权利要求1所述的方法，其特征在于，根据所述目标信息，对所述原始高分辨率遥感图像进行自适应的图像分区，得到不同的簇侯检区，进而得到必检区和非必检区，包括：
    S21：根据目标和目标的位置进行聚类，计算聚类的类别数的初始值，其中所述目标信息包括目标的位置、尺寸和类别；
    S22：根据所述类别数的初始值，执行聚类算法，得到对应数量的簇；
    S23：对每个簇构建一个外接矩形，计算该矩形的边长以及矩形内目标的密度；
    S24：如果矩形的边长超过边长阈值且矩形内目标的密度超过密度阈值，更新所述类别数的值，返回步骤S22重新聚类；反之则根据目标的密度和尺寸，选择能够处理所述密度和尺寸的模型集合；
    S25：结合目标的运动信息，计算出簇内目标可能出现的区域，设为簇侯检区，所有簇侯检区共同组成必检区，所述原始高分辨率遥感图像中不为必检区的区域则为非必检区，其中所述目标的运动信息存储在所述计算平台上。
  5. 根据权利要求4所述的方法，其特征在于，进行必检区的模型选择，生成必检区的目标检测方案，包括：
    S31：根据目标的密度和尺寸，选择能够处理所述密度和尺寸的模型集合；
    S32：根据所述模型集合中模型的不同输入尺寸将其中一个簇侯检区均匀分割，产生对该簇侯检区可行的至少一种模型组合；
    S33：对所述必检区中的每一个簇侯检区，执行S32，将所有簇侯检区的模型组合之后形成至少一种模型方案；
    S34：根据每个模型方案的精度和延迟信息，计算每个模型方案对于检测所述必检区的精度和总延迟；
    S35：在所有模型方案中以整体精度优先方式选择总延迟满足所述计算平台的延迟限制的方案作为必检区的目标检测方案。
  6. 一种高分辨率遥感图像的目标检测装置，其特征在于，应用于计算平台上，包括：
    第一获取模块，用于从传感器获取原始高分辨率遥感图像；
    第二获取模块，用于从历史检测结果中获取所述原始高分辨率遥感图像所对应区域以及所述区域向四周扩展预定距离的扩展区域内的目标信息；
    分区模块，用于根据所述目标信息，对所述原始高分辨率遥感图像进行自适应的图像分区，得到不同的簇侯检区，进而得到必检区和非必检区；
    生成模块，用于进行必检区的模型选择，生成必检区的目标检测方案；
    执行模块，用于执行所述目标检测方案，得到各必检区的检测结果；
    检测模块，用于判断计算平台是否有多余计算资源以检测非必检区，若有，则进行非必检区的动态分区检测，得到非必检区的检测结果，将所述各必检区的检测结果和所述非必检区的检测结果合并为目标检测结果，否则将所述各必检区的检测结果作为目标检测结果。
  7. 一种电子设备，其特征在于，包括：
    一个或多个处理器；
    存储器，用于存储一个或多个程序；
    当所述一个或多个程序被所述一个或多个处理器执行，使得所述一个或多个处理器实现如权利要求1-5任一项所述的高分辨率遥感图像的目标检测方法。
  8. 一种计算机可读存储介质，其上存储有计算机指令，其特征在于，该指令被处理器执行时实现如权利要求1-5中任一项所述的高分辨率遥感图像的目标检测方法的步骤。
PCT/CN2023/088349 2022-05-05 2023-04-14 一种高分辨率遥感图像的目标检测方法及装置 Ceased WO2023213178A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US18/464,249 US12014536B2 (en) 2022-05-05 2023-09-10 Target detection method and device for high-resolution remote sensing image

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202210480814.4A CN114581781B (zh) 2022-05-05 2022-05-05 一种高分辨率遥感图像的目标检测方法及装置
CN202210480814.4 2022-05-05

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US18/464,249 Continuation US12014536B2 (en) 2022-05-05 2023-09-10 Target detection method and device for high-resolution remote sensing image

Publications (1)

Publication Number Publication Date
WO2023213178A1 true WO2023213178A1 (zh) 2023-11-09

Family

ID=81783663

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2023/088349 Ceased WO2023213178A1 (zh) 2022-05-05 2023-04-14 一种高分辨率遥感图像的目标检测方法及装置

Country Status (3)

Country Link
US (1) US12014536B2 (zh)
CN (1) CN114581781B (zh)
WO (1) WO2023213178A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117612008A (zh) * 2023-11-29 2024-02-27 广东技术师范大学 一种基于高光谱识别技术的盐湖水体检测方法与装置
CN119045389A (zh) * 2024-10-29 2024-11-29 四川伟博震源科技有限公司 一种建筑施工安全控制系统和方法

Families Citing this family (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114581781B (zh) * 2022-05-05 2022-08-09 之江实验室 一种高分辨率遥感图像的目标检测方法及装置
CN117994506B (zh) * 2024-04-07 2024-08-20 厦门大学 一种动态知识集成的遥感图像显著性目标检测方法
CN118658070B (zh) * 2024-08-20 2024-11-19 浙江省城乡规划设计研究院 基于遥感影像的连片低效工业用地变化检测方法及系统
CN119596356B (zh) * 2024-11-15 2025-09-30 哈尔滨工业大学 一种基于超宽幅旋扫遥感图像的高价值目标快速检测与定位方法及其系统
CN120029855B (zh) * 2025-04-18 2025-10-21 北京帮威客科技有限公司 一种基于边缘计算的遥感监测方法及系统

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102043958A (zh) * 2010-11-26 2011-05-04 华中科技大学 一种高分辨率遥感图像多类目标检测识别方法
US20190273910A1 (en) * 2018-03-05 2019-09-05 Valve Corporation Controlling image display via real-time compression in peripheral image regions
CN112381053A (zh) * 2020-12-01 2021-02-19 连云港豪瑞生物技术有限公司 具有图像跟踪功能的环保监测系统
CN113239815A (zh) * 2021-05-17 2021-08-10 广东工业大学 一种基于真实语义全网络学习的遥感影像分类方法、装置及设备
CN114581781A (zh) * 2022-05-05 2022-06-03 之江实验室 一种高分辨率遥感图像的目标检测方法及装置

Family Cites Families (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
TWI492166B (zh) * 2012-01-12 2015-07-11 Kofax Inc 行動影像擷取和處理的系統和方法
CN109416728A (zh) * 2016-09-30 2019-03-01 富士通株式会社 目标检测方法、装置以及计算机系统
CN108445480B (zh) * 2018-02-02 2022-05-03 重庆邮电大学 基于激光雷达的移动平台自适应扩展目标跟踪系统及方法
CN108337399A (zh) * 2018-02-27 2018-07-27 贵州省第三测绘院 遥感图像处理方法和装置
CN110390292B (zh) * 2018-10-26 2022-03-04 西安电子科技大学 基于动态关联模型的遥感视频车辆目标检测与跟踪方法
CN109948415A (zh) * 2018-12-30 2019-06-28 中国科学院软件研究所 基于背景过滤和尺度预测的光学遥感图像目标检测方法
CN109902627A (zh) * 2019-02-28 2019-06-18 中科创达软件股份有限公司 一种目标检测方法及装置
CN112199984B (zh) * 2020-07-10 2023-05-12 北京理工大学 一种大尺度遥感图像的目标快速检测方法
CN112132093B (zh) * 2020-09-30 2024-01-05 湖南省气象科学研究所 高分辨率遥感图像目标检测方法、装置和计算机设备
CN112668390B (zh) * 2020-11-17 2024-08-13 福建省星云大数据应用服务有限公司 一种高效率的单幅遥感影像目标检测方法及系统
CN113971653A (zh) * 2021-04-06 2022-01-25 国家气象中心 面向遥感图像的目标检测方法、装置、设备及存储介质
CN113705532B (zh) * 2021-09-10 2023-05-23 中国人民解放军国防科技大学 基于中低分辨率遥感图像的目标检测方法、装置及设备
CN114244904B (zh) * 2021-12-15 2023-05-09 平安科技（深圳）有限公司 遥感数据传输方法、装置、设备及介质
CN115004245A (zh) * 2021-12-17 2022-09-02 商汤国际私人有限公司 目标检测方法、装置、电子设备和计算机存储介质

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102043958A (zh) * 2010-11-26 2011-05-04 华中科技大学 一种高分辨率遥感图像多类目标检测识别方法
US20190273910A1 (en) * 2018-03-05 2019-09-05 Valve Corporation Controlling image display via real-time compression in peripheral image regions
CN112381053A (zh) * 2020-12-01 2021-02-19 连云港豪瑞生物技术有限公司 具有图像跟踪功能的环保监测系统
CN113239815A (zh) * 2021-05-17 2021-08-10 广东工业大学 一种基于真实语义全网络学习的遥感影像分类方法、装置及设备
CN114581781A (zh) * 2022-05-05 2022-06-03 之江实验室 一种高分辨率遥感图像的目标检测方法及装置

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117612008A (zh) * 2023-11-29 2024-02-27 广东技术师范大学 一种基于高光谱识别技术的盐湖水体检测方法与装置
CN119045389A (zh) * 2024-10-29 2024-11-29 四川伟博震源科技有限公司 一种建筑施工安全控制系统和方法

Also Published As

Publication number Publication date
US12014536B2 (en) 2024-06-18
CN114581781A (zh) 2022-06-03
CN114581781B (zh) 2022-08-09
US20230419639A1 (en) 2023-12-28

Similar Documents

Publication Publication Date Title
CN114581781B (zh) 一种高分辨率遥感图像的目标检测方法及装置
EP3757899B1 (en) Neural architecture for self supervised event learning and anomaly detection
CN114902190B (zh) 用于现场故障检测的物理操作参数的硬件控制更新
Liu et al. Masc: Multi-scale affinity with sparse convolution for 3d instance segmentation
Liu et al. Embedded online fish detection and tracking system via YOLOv3 and parallel correlation filter
US20190114546A1 (en) Refining labeling of time-associated data
Blecic et al. Cellular automata simulation of urban dynamics through GPGPU
CN112991537B (zh) 城市场景重建方法、装置、计算机设备和存储介质
JP2023516500A (ja) 画像ベースの場所決定のためのシステム及び方法
US11875523B2 (en) Adaptive stereo matching optimization method and apparatus, device and storage medium
US20190163978A1 (en) Budget-aware method for detecting activity in video
CN113076924A (zh) 基于无人船的河面垃圾实时检测方法、装置及存储介质
KR102319145B1 (ko) 고해상도 해양 데이터를 생성하기 위한 방법 및 이를 이용한 장치
Scalcon et al. Ai-powered video monitoring: Assessing the nvidia jetson orin devices for edge computing applications
Wu et al. Gauspu: 3d gaussian splatting processor for real-time slam systems
CN115375887A (zh) 一种移动目标轨迹预测方法、装置、设备及介质
CN117496321A (zh) 一种基于两种模态图像融合的目标检测方法
Wang et al. An Improved Object Detection Method for Underwater Sonar Image Based on PP‐YOLOv2
Zhang et al. DSP-based traffic target detection for intelligent transportation
CN101833760A (zh) 一种基于图像块的背景建模方法及背景建模装置
CN113313171A (zh) 基于fpga实现图像识别的方法、装置、设备及存储介质
Liu et al. Target classification and recognition for high-resolution remote sensing images: using the parallel cross-model neural cognitive computing algorithm
CN112101538B (zh) 基于内存计算的图神经网络硬件计算系统及方法
CN120219296A (zh) 变电站瓷套类套管破损的检测方法、装置、计算机设备、存储介质和程序产品
CN118968240A (zh) 一种红外图像目标检测方法、装置、设备及介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 23799165

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 23799165

Country of ref document: EP

Kind code of ref document: A1

122 Ep: pct application non-entry in european phase

Ref document number: 23799165

Country of ref document: EP

Kind code of ref document: A1

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 17/06/2025)