Large scene target processing method and processing system based on image features

By setting multiple scale reduction levels in large scene images and determining the optimal scale, extracting and correlating target features, the problem of high computing resource consumption in high-resolution image processing is solved, and efficient and accurate object detection is achieved.

CN118297795BActive Publication Date: 2024-11-05SICHUAN NATIONAL INNOVATION VISION UHD VIDEO TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410487154.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-04-23
Publication Date
2024-11-05
Estimated Expiration
2044-04-23

AI Technical Summary

Technical Problem

When processing high-resolution images in large scenarios, it is difficult for the prior art to reduce computing resource consumption while retaining target information. Traditional methods cannot effectively ensure the object detection effect by reducing the image size, and the calculation complexity is still high.

Method used

By acquiring large scene images, setting multiple reduction ratio levels, extracting target features (number, size, bounding box overlap rate), comprehensive comparison to determine the optimal reduction ratio, reducing the image and correlating the target in the original image, using neural network model Perform target recognition.

Benefits of technology

While retaining sufficient target information, reducing computing complexity, improving the efficiency and accuracy of large-scenario target processing, and reducing computing resource consumption.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118297795B_ABST
    Figure CN118297795B_ABST
Patent Text Reader

Abstract

The present invention provides a large scene target processing method and a processing system thereof based on image features, belonging to the field of image processing technology, comprising obtaining an original image in a large scene, and setting a reduction ratio level according to the size of the original image; extracting target features in the original image processed at each reduction ratio level, wherein the target features include the number of targets, the target size and the overlap rate of the target boundary box; comprehensively comparing the target features at each reduction ratio level, and determining the best reduction ratio according to the comparison result; reducing the original image according to the best reduction ratio to obtain an image with the best reduction ratio, and associating the target with the original image in the image with the best reduction ratio. By adopting the above technical solution, while retaining sufficient target information, the calculation complexity can be effectively reduced, and the efficiency and accuracy of large scene target processing can be improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a large scene target processing method based on image features and a processing system thereof. Background Art

[0002] With the rapid development of image processing technology, the demand for high-resolution image processing in large scenes is increasing. Especially in large dynamic scenes, such as natural environments covering thousands of square kilometers or social scenes involving thousands to tens of thousands of people, the resolution of images often reaches the billion-pixel level. Such images contain rich multi-scale information, which poses great challenges to tasks such as target detection, recognition and tracking.

[0003] However, high-pixel image processing is often accompanied by huge computing resource consumption, which places extremely high demands on existing computing devices. Therefore, how to reduce the amount of data to be processed while retaining the original image information has become an urgent problem to be solved in the field of large-scene object processing.

[0004] Traditional image processing methods usually reduce the computational complexity by directly reducing the size of the image. However, this method often cannot ensure that all objects in the original image can be effectively detected in the subsequent steps. The choice of reduction ratio has an important impact on the effect of object detection. If the reduction ratio is too large, the target information may be lost; if the reduction ratio is too small, although more target information is retained, the computational complexity is still high.

[0005] Therefore, it is necessary to provide a large scene target processing method and a processing system based on image features to solve the above technical problems. Summary of the invention

[0006] To solve the above technical problems, the present invention provides a large-scene target processing method and a processing system based on image features, which can reduce the consumption of computing resources while ensuring the target detection effect, and provide a more efficient and reliable solution for practical applications.

[0007] The present invention provides a large scene target processing method based on image features, and the processing method comprises the following steps:

[0008] Get the original image of the large scene and set the reduction ratio level according to the original image size;

[0009] Extracting target features in the original image after being processed at each reduction scale level, wherein the target features include the number of targets, the size of targets, and the overlap rate of target bounding boxes;

[0010] Comprehensively compare the target features at each reduction ratio level, and determine the best reduction ratio based on the comparison results;

[0011] The original image is reduced according to the optimal reduction ratio to obtain an optimal reduction ratio image, and the target is associated with the original image in the optimal reduction ratio image.

[0012] Preferably, the obtaining of the original image in the large scene and setting the reduction ratio level according to the size of the original image comprises the following steps:

[0013] Obtaining the original image size, wherein the original image size includes the height and width of the original image;

[0014] Dividing the size lower than the original image size based on a preset size interval to obtain multiple size levels;

[0015] The original image size is divided by a plurality of size levels to obtain a plurality of reduction ratio levels corresponding to the plurality of size levels.

[0016] Preferably, the step of extracting target features from the original image after being processed at each reduction scale level, wherein the target features include the number of targets, the size of targets, and the overlap rate of target bounding boxes, comprises the following steps:

[0017] The original image is reduced according to each reduction ratio level to obtain a reduced image corresponding to each reduction ratio level;

[0018] At each reduction ratio level, an object detection algorithm is used to extract object features in each reduced image, wherein the object features include the number of objects contained in each reduced image, the object size, and the overlap rate of the object bounding box.

[0019] Preferably, the step of comprehensively comparing the target features at each reduction ratio level and determining the best reduction ratio according to the comparison results comprises the following steps:

[0020] Obtaining current application requirements, and sorting the number of targets, target sizes, and overlapping rates of target bounding boxes based on the current application requirements to obtain a comparison order;

[0021] Based on a preset threshold, and in accordance with the comparison order, the number of objects, the size of objects, and the overlap rate of the object bounding boxes at each reduction ratio level are compared in sequence, wherein the preset threshold includes a number threshold, a size threshold, and an overlap rate threshold corresponding to the number of objects, the size of objects, and the overlap rate of the object bounding boxes, respectively;

[0022] Obtaining at least one pair of reduction ratio levels in which the number of objects, the size of objects, and the overlap rate of the object bounding boxes can respectively satisfy a number threshold, a size threshold, and an overlap rate threshold;

[0023] Selecting a pair of highest reduction ratio levels from at least one pair of reduction ratio levels, wherein the higher the reduction ratio level, the greater the reduction ratio, and the less information the corresponding reduced image contains;

[0024] The higher reduction ratio level of a pair of highest-ranked reduction ratio levels is selected as the optimal reduction ratio.

[0025] Preferably, the current application requirement is a purpose of target detection, wherein the purpose of target detection includes counting requirements, location requirements and measurement requirements.

[0026] Preferably, if the current application requirement is a counting requirement, the comparison order is the number of targets, the target size, and the overlap rate of the target bounding boxes;

[0027] If the current application requirement is a location requirement, the comparison order is the overlap rate of the target bounding box, the target size, and the number of targets;

[0028] If the current application requirement is a measurement requirement, the comparison order is target size, target bounding box overlap rate, and target quantity.

[0029] Preferably, the step of reducing the original image according to the optimal reduction ratio to obtain the optimal reduction ratio image, and associating the target with the original image in the optimal reduction ratio image comprises:

[0030] Reducing the original image according to an optimal reduction ratio to obtain an image to be processed, wherein the image to be processed includes size information;

[0031] Based on the size information of the image to be processed, the image to be processed is input into a corresponding target area search model, and at least one target area is determined according to a generation result of the target area search model, wherein the target area search model is a neural network model trained by a set training method, and the input for training the target area search model is image training samples at different size levels;

[0032] Input each target area into the target identifier for position detection, and determine the target position frame in each target area;

[0033] According to the size information and the best reduction ratio of each target area, the corresponding target position box is associated with the original image.

[0034] The present invention also provides a large scene target processing system based on image features, which is applied to the above target processing method. The processing system includes:

[0035] A reduction ratio level determination module is used to obtain the original image in a large scene and set the reduction ratio level according to the size of the original image;

[0036] A target feature extraction module, used to extract target features in the original image after being processed at each reduction scale level, wherein the target features include the number of targets, the size of targets and the overlap rate of target bounding boxes;

[0037] A comprehensive comparison module is used to comprehensively compare the target features at each reduction ratio level and determine the best reduction ratio based on the comparison results;

[0038] The target association module is used to reduce the original image according to the optimal reduction ratio to obtain the optimal reduction ratio image, and associate the target with the original image in the optimal reduction ratio image.

[0039] Compared with the related art, the large scene target processing method and processing system based on image features provided by the present invention have the following beneficial effects:

[0040] The present invention obtains the original image in a large scene, and sets the reduction ratio level according to the size of the original image; extracts the target features in the original image processed by each reduction ratio level, wherein the target features include the number of targets, the target size and the overlap rate of the target boundary box; comprehensively compares the target features at each reduction ratio level, and determines the best reduction ratio according to the comparison result; reduces the original image according to the best reduction ratio to obtain the best reduction ratio image, and associates the target to the original image in the best reduction ratio image. By adopting the above technical solution, the original image can be processed at different reduction ratios, the target features at each reduction ratio can be extracted, and the best reduction ratio can be determined by comprehensive comparison, so that while retaining sufficient target information, the computational complexity can be effectively reduced, and the efficiency and accuracy of large scene target processing can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 A flowchart of a large scene target processing method based on image features provided by the present invention;

[0042] Figure 2 A flowchart of step S1 of a large scene target processing method based on image features provided by the present invention;

[0043] Figure 3 A flowchart of step S2 of a large scene target processing method based on image features provided by the present invention;

[0044] Figure 4 A flowchart of step S3 of a large scene target processing method based on image features provided by the present invention;

[0045] Figure 5 A flowchart of step S4 of a large scene target processing method based on image features provided by the present invention;

[0046] Figure 6 A schematic diagram of the module structure of a large scene target processing system based on image features provided by the present invention. DETAILED DESCRIPTION

[0047] The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the present invention, rather than to limit the present invention. It should also be noted that, for ease of description, only the parts related to the present invention, rather than all structures, are shown in the accompanying drawings. In addition, the embodiments of the present invention and the features in the embodiments may be combined with each other without conflict.

[0048] It should also be noted that, for ease of description, only the part relevant to the present invention but not all the content is shown in the accompanying drawings. It should be mentioned before discussing the exemplary embodiments in more detail that some exemplary embodiments are described as the processing or method described as a flow chart. Although the flow chart describes each operation (or step) as being processed sequentially, many operations therein can be implemented in parallel, concurrently or simultaneously. In addition, the order of each operation can be rearranged. When its operation is completed, the processing can be terminated, but it can also have additional steps not included in the accompanying drawings. The processing can correspond to a method, a function, a procedure, a subroutine, a subprogram, etc.

[0049] Embodiment 1

[0050] The present invention provides a large scene target processing method based on image features, referring to Figure 1 As shown, the processing method includes the following steps:

[0051] S1: Obtain the original image of the large scene and set the reduction ratio level according to the size of the original image.

[0052] In this embodiment, the original image of the large scene is first obtained. The original image may be an image corresponding to a video frame in a billion-pixel video collected from a dynamic large scene. The large scene may contain multiple dynamic targets with extremely high resolution and rich information content, such as a natural environment covering thousands of square kilometers or a social scene containing thousands to tens of thousands of people at the same time.

[0053] Next, considering that the original image may contain a huge amount of pixels, direct processing will consume a lot of computing resources, so it is necessary to set a series of reduction ratio levels according to the size of the original image, and then a series of reduction ratio levels can be set according to the size of the original image. Specifically, these reduction ratio levels are determined according to the size smaller than the original image size and the preset size spacing. For example, multiple different reduction ratios can be set, such as 0.5, 0.25, 0.1, etc., so that the appropriate reduction ratio can be selected according to the needs in subsequent processing.

[0054] By setting different reduction ratio levels, the amount of data for subsequent processing can be effectively reduced while retaining the key information of the original image, thereby improving processing efficiency and reducing computing costs.

[0055] S2: extracting target features from the original image after being processed at each reduction scale level, wherein the target features include the number of targets, the size of targets, and the overlap rate of target bounding boxes.

[0056] In this embodiment, for each image at each reduction ratio level, it is necessary to extract the target features therein. These target features may include the number of targets, the size of targets, and the overlap rate of target bounding boxes. Specifically, the target detection algorithm can be used to identify the number of targets in the image (such as the number of people and vehicles in the scene), and calculate the size of each target (such as the pixel area and length of the target) and the bounding box. At the same time, it is also necessary to calculate the overlap rate between different target bounding boxes (for evaluating the degree of occlusion between targets) to understand the relative position and relationship between the targets. Finally, by obtaining detailed information about the target features in the images at different reduction ratio levels, data support is provided for subsequent comparison and analysis.

[0057] S3: Comprehensively compare the target features at each reduction ratio level, and determine the optimal reduction ratio based on the comparison results.

[0058] In this embodiment, a comprehensive analysis is performed on the target features extracted in the previous step. Specifically, it is necessary to evaluate the difference in the number of targets at different reduction ratio levels, the change in target size, and the fluctuation in the target bounding box overlap rate. By comparing these features, the best reduction ratio that can retain sufficient information and reduce the amount of calculation can be found. This ratio should be able to keep the target features relatively stable while reducing the amount of image data as much as possible, providing an optimization solution for subsequent image processing work.

[0059] S4: reducing the original image according to the optimal reduction ratio to obtain an optimal reduction ratio image, and associating the target with the original image in the optimal reduction ratio image.

[0060] In this embodiment, after determining the optimal reduction ratio, the original image is reduced according to the optimal reduction ratio to obtain the optimal reduction ratio image. This optimal reduction ratio image will serve as the main object of subsequent processing and analysis because it not only retains the key information of the original image, but also greatly reduces the processing difficulty and computing cost.

[0061] After obtaining the image with the best reduction ratio, it is also necessary to associate the target in the image with the original image. In this way, even in the reduced image, the target position corresponding to the original image can still be accurately found, which facilitates subsequent target positioning, recognition and analysis.

[0062] Specifically, refer to Figure 2 As shown, step S1 includes the following steps:

[0063] S101: Acquire the original image size, wherein the original image size includes the height and width of the original image.

[0064] In this embodiment, the size information of the original image must first be obtained, which is the basis for subsequent image processing. The size of the original image usually includes two dimensions, height and width, and is measured in pixels. For example, an original image may have a height of 8000 pixels and a width of 6000 pixels.

[0065] Specifically, the original image size can be directly read through the image processing software or the relevant functions in the programming library. The acquisition of the original image size provides an accurate size reference for subsequent image processing, so that the subsequent processing process can be optimized for the specific image size.

[0066] S102: Dividing the size lower than the original image size based on a preset size interval to obtain a plurality of size levels.

[0067] In this embodiment, after the original image size is obtained, a series of size levels lower than the original image size are divided according to a preset size spacing. The preset size spacing can be set according to actual needs, for example, it can be 100 pixels, 500 pixels, etc.

[0068] Based on this spacing, starting from the original image size, the size is gradually reduced to obtain multiple different size levels. These size levels will serve as possible reduction targets in subsequent image processing to meet the processing requirements in different scenarios.

[0069] For example, if the original image size is 8000*6000 pixels and the preset size interval is 500 pixels, then a series of size levels can be obtained, such as 7500*5500 pixels, 7000*5000 pixels, etc. By dividing into multiple size levels, more options are provided for subsequent image reduction, which can more flexibly adapt to different processing requirements.

[0070] S103: Dividing the original image size by a plurality of size levels to obtain a plurality of reduction ratio levels corresponding to the plurality of size levels.

[0071] In this embodiment, after obtaining multiple size levels, the reduction ratio of each size level relative to the original image is calculated. Specifically, this can be achieved by dividing the original image size by each size level. The calculated reduction ratio will serve as a reference for reducing the image in subsequent image processing.

[0072] For example, if the original image size is 8000*6000 pixels and a certain size level is 7000*5000 pixels, then the corresponding reduction ratio is (8000 / 7000)*(6000 / 5000), which is about 1.14x1.2. Therefore, multiple reduction ratio levels corresponding to multiple size levels can be obtained.

[0073] By calculating the reduction ratio level, a specific quantitative index is provided for the subsequent image reduction, which enables more precise control of the image reduction degree, thereby reducing the processing difficulty and calculation cost while ensuring image quality.

[0074] Specifically, refer to Figure 3 As shown, step S2 includes the following steps:

[0075] S201: performing reduction processing on the original image according to each reduction ratio level to obtain a reduced image corresponding to each reduction ratio level.

[0076] In this embodiment, the original image is reduced according to a preset reduction ratio level, the purpose of which is to reduce the computational complexity of subsequent target feature extraction while retaining sufficient information to support the accuracy of the target detection algorithm.

[0077] The specific process of the reduction processing includes: determining the reduction ratio level, and then using the image scaling algorithm to proportionally reduce the original image to obtain a reduced image corresponding to each reduction ratio level, wherein the image scaling algorithm includes but is not limited to bilinear interpolation or bicubic interpolation.

[0078] Through the reduction process, a series of reduced images of different sizes can be obtained. These images retain the key information of the original image while reducing the complexity of data processing, which facilitates the subsequent target feature extraction. At the same time, since the reduced image still retains the key information of the original image, it will not have a significant impact on the accuracy of the target detection algorithm.

[0079] S202: At each reduction ratio level, extract target features in each reduced image using a target detection algorithm, wherein the target features include the number of targets contained in each reduced image, the target size, and the overlap rate of the target bounding box.

[0080] In this embodiment, a target detection algorithm is used to extract target features in the reduced image. The target detection algorithm specifically includes:

[0081] Candidate region generation: First, a series of candidate regions that may contain targets are generated in the reduced image through methods such as sliding windows or region proposal networks (RPNs). These candidate regions may cover different positions and scales in the image to capture all potential targets as much as possible.

[0082] Feature extraction: For each candidate region, a deep convolutional neural network (CNN) is used to extract features. CNN extracts high-level semantic features from the candidate region through multi-layer convolution, pooling, and activation functions. These features are crucial for distinguishing between targets and backgrounds.

[0083] Classification and localization: The extracted features are sent to the classifier and bounding box regressor for processing. The classifier is responsible for determining whether each candidate region contains the target, while the bounding box regressor is responsible for adjusting the bounding box of the candidate region to make it more accurately fit the actual position of the target.

[0084] Non-maximum suppression (NMS): Since multiple overlapping bounding boxes will be generated, non-maximum suppression is required to eliminate redundant detection results. NMS compares the scores and overlaps of different bounding boxes, retains the bounding box with the highest score, and suppresses other bounding boxes with high overlaps.

[0085] Feature output: After the above steps, the target position of each target (represented by a bounding box), the number of targets, and the overlap rate of the bounding boxes are output.

[0086] Specifically, refer to Figure 4 As shown, step S3 includes the following steps:

[0087] S301: Obtain current application requirements, and sort the number of objects, object sizes, and overlap rates of object bounding boxes based on the current application requirements to obtain a comparison order.

[0088] In this embodiment, the specific requirements of the current application are first clarified, that is, the purpose of target detection. For example, if the purpose of target detection is to count people, then the accuracy of the number of targets may be the most critical; if it is to locate the target, then the accuracy and clarity of the target bounding box may be more important; and if it is to measure the size of the target, then the accuracy of the target size becomes a key indicator.

[0089] Based on these application requirements, the number of targets, target size, and target bounding box overlap are sorted. If the target detection purpose is based on counting requirements, the number of targets is ranked first, followed by the target size, and finally the target bounding box overlap. This sorting ensures that the most critical indicator, the accuracy of the number of targets, is met first in the subsequent comparison process.

[0090] Alternatively, if the purpose of object detection is location requirement, the comparison order is the overlap rate of the object bounding box, the object size and the number of objects;

[0091] Alternatively, if the purpose of target detection is measurement requirements, the comparison order is target size, target bounding box overlap, and target number.

[0092] By sorting according to application requirements, it is possible to more specifically compare target features at different reduction ratio levels, thereby finding the reduction ratio that best meets the application requirements.

[0093] S302: Based on a preset threshold and in accordance with the comparison order, the number of targets, the size of targets and the overlap rate of target bounding boxes at each reduction ratio level are compared in sequence, wherein the preset threshold includes a number threshold, a size threshold and an overlap rate threshold corresponding to the number of targets, the size of targets and the overlap rate of target bounding boxes, respectively.

[0094] In this embodiment, a series of preset thresholds are set, including a quantity threshold, a size threshold and an overlap rate threshold. These thresholds are determined according to actual application requirements and are used to evaluate the satisfaction degree of target features at different reduction ratio levels.

[0095] Then, according to the comparison order determined in step S301, the number of objects, the size of objects, and the overlap rate of the object bounding boxes at each reduction ratio level are compared in turn. For example, if the current application requirement is a counting requirement, firstly, whether the number of objects at each reduction ratio level meets the number threshold, then whether the size of the object meets the size threshold, and finally whether the overlap rate of the object bounding boxes meets the overlap rate threshold.

[0096] And so on for location needs and measurement needs.

[0097] By setting the threshold and making comparisons in order, we can objectively and accurately evaluate the satisfaction degree of target features at different reduction ratio levels, and provide a basis for the subsequent selection of the best reduction ratio.

[0098] S303: Obtain at least one pair of reduction ratio levels in which the number of objects, the size of objects, and the overlap rate of the object bounding boxes can respectively satisfy a number threshold, a size threshold, and an overlap rate threshold.

[0099] In this embodiment, those reduction ratio levels that can meet the preset threshold are screened out. These levels meet the basic requirements of the application needs in terms of the number of objects, object size, and overlap rate of the object bounding box, and those reduction ratio levels that do not meet the requirements are excluded, thereby narrowing the scope of subsequent selection and improving the efficiency of selection.

[0100] At the same time, in multiple one-to-one comparison processes, due to the complexity of image scaling, there may be multiple pairs of two reduction ratio levels that can meet these threshold conditions. This is because different reduction ratios may reduce the amount of computation to different degrees while retaining key information. Therefore, one or more pairs of reduction ratio levels may be obtained, all of which can achieve different degrees of reduction in computation while meeting application requirements.

[0101] S304: Selecting a pair of highest reduction ratio levels from at least one pair of reduction ratio levels, wherein a higher reduction ratio level means a larger reduction ratio, and a corresponding reduced image contains less information.

[0102] In this embodiment, a pair with the highest level is selected from the reduction ratio levels that meet a preset threshold. The level here is determined according to the reduction ratio. The larger the reduction ratio, the higher the level. The pair with the highest level is selected because while keeping the target features to meet application requirements, a larger reduction ratio means less calculation and higher processing efficiency.

[0103] S305: Selecting a higher reduction ratio level from a pair of highest-level reduction ratio levels as the optimal reduction ratio.

[0104] In this embodiment, the one with the highest reduction ratio is selected from a pair of the highest reduction ratio levels as the best reduction ratio. This is because, when the application requirements are equally met, a larger reduction ratio means less calculation and higher processing efficiency.

[0105] Specifically, refer to Figure 5 As shown, step S4 includes the following steps:

[0106] S401: reducing the original image according to an optimal reduction ratio to obtain an image to be processed, wherein the image to be processed includes size information.

[0107] In this embodiment, the original image is scaled according to the determined optimal reduction ratio, and a to-be-processed image is obtained through the reduction process. The size of this image is smaller than the original image, but sufficient key information is retained for subsequent analysis and processing. By precisely controlling the reduction ratio, the integrity of the image information can be maintained while significantly reducing the amount of data for subsequent processing, thereby improving processing efficiency and reducing computing costs.

[0108] S402: Based on the size information of the image to be processed, the image to be processed is input into a corresponding target area search model, and at least one target area is determined according to the generation result of the target area search model, wherein the target area search model is a neural network model trained by a set training method, and the input for training the target area search model is image training samples at different size levels.

[0109] In this embodiment, based on the size information of the image to be processed, it is input into a pre-trained target area search model, and the size of the image training samples used in the training of the target area search model corresponds to the size information of the image to be processed. This model is a neural network model optimized by a specific training method, which can automatically adjust the processing strategy according to the input image size to accurately identify the target area in the image.

[0110] During the model training phase, we used image training samples of different size levels to ensure that the model can adapt to input images of various sizes. In this way, no matter how the size of the processed image changes, the model can maintain stable performance and output accurate target area information.

[0111] For example, the target region search model may be a convolutional neural network (CNN) model, which has been trained with a large amount of data and learned to extract features from images and identify target regions. When the image to be processed is input to the model, the model automatically analyzes the image content and outputs the coordinates and size information of the target region.

[0112] Finally, by utilizing the pre-trained target region search model, the target region can be quickly and accurately identified from the image to be processed, providing a reliable foundation for subsequent target position detection.

[0113] S403: Input each target area into the target identifier for position detection, and determine the target position frame of each target area.

[0114] In this embodiment, each target region output from the target region search model is input into a target identifier, which is a model specifically used to detect the position of a target in an image, and can perform detailed analysis and processing on each target region to determine the precise position of the target in the region.

[0115] For example, the object recognizer is a neural network model trained using the YOLO (You Only Look Once) algorithm. After being trained with a large amount of labeled data, it has learned how to accurately detect the location of objects in an image. When the target area is input to the recognizer, it outputs the location box information of each target, including the coordinate values ​​of the upper left corner and the lower right corner.

[0116] S404: according to the size information and the best reduction ratio of each target area, the corresponding target position frame is associated with the original image.

[0117] In this embodiment, the target position frame output by the target identifier is associated back to the original image according to the size information of the target area and the optimal reduction ratio. Since the size of the original image is much larger than the image to be processed, the target position frame needs to be enlarged accordingly according to the reduction ratio to ensure that it can accurately correspond to the target position in the original image.

[0118] Specifically, the corresponding size of the target position box in the original image can be obtained by calculating the product of the reduction ratio and the target area size. Then, according to the coordinate mapping relationship between the image to be processed and the original image, the coordinate value of the target position box is converted back to the coordinate system of the original image. Then, the position of the target is accurately marked in the original image, thereby realizing target processing of the original image collected in a large scene.

[0119] The working principle of a large scene target processing method based on image features provided by the present invention is as follows: the present invention obtains the original image in the large scene, and sets the reduction ratio level according to the size of the original image; extracts the target features in the original image processed by each reduction ratio level, wherein the target features include the number of targets, the target size and the overlap rate of the target bounding box; comprehensively compares the target features under each reduction ratio level, and determines the best reduction ratio according to the comparison results; reduces the original image according to the best reduction ratio to obtain the best reduction ratio image, and associates the target to the original image in the best reduction ratio image. By adopting the above technical solution, the original image can be processed at different reduction ratios, the target features under each reduction ratio can be extracted, and the best reduction ratio can be determined by comprehensive comparison, so that while retaining sufficient target information, the computational complexity can be effectively reduced, and the efficiency and accuracy of large scene target processing can be improved.

[0120] Embodiment 2

[0121] The present invention also provides a large scene target processing system based on image features, which is applied to the above target processing method. The processing system includes:

[0122] The reduction ratio level determination module 100 is used to obtain the original image in a large scene and set the reduction ratio level according to the size of the original image.

[0123] The target feature extraction module 200 is used to extract target features in the original image after being processed at each reduction scale level, wherein the target features include the number of targets, the size of targets and the overlap rate of target bounding boxes.

[0124] The comprehensive comparison module 300 is used to comprehensively compare the target features at each reduction ratio level and determine the best reduction ratio according to the comparison results.

[0125] The target associating module 400 is used to reduce the original image according to the optimal reduction ratio to obtain the optimal reduction ratio image, and associate the target with the original image in the optimal reduction ratio image.

[0126] The present application is described with reference to flowcharts and / or block diagrams of methods, devices (systems) and computer program products according to embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0127] A person skilled in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium, and the storage medium includes a read-only memory (ROM), a random access memory (RAM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), a one-time programmable read-only memory (OTPROM), an electronically erasable rewritable read-only memory (EEPROM), a compact disc (CD-ROM) or other optical disc storage, magnetic disk storage, magnetic tape storage, or any other computer-readable medium that can be used to carry or store data.

[0128] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device that includes a series of elements includes not only those elements, but also other elements that are not explicitly listed, or includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device that includes the elements.

Claims

1. A large scene target processing method based on image features, characterized in that: The treatment method includes the following steps: Get the original image of the large scene and set the reduction ratio level according to the original image size; Extracting target features in the original image after being processed at each reduction scale level, wherein the target features include the number of targets, the size of targets, and the overlap rate of target bounding boxes; Comprehensively compare the target features at each reduction ratio level, and determine the best reduction ratio based on the comparison results; Reducing the original image according to the optimal reduction ratio to obtain an optimal reduction ratio image, and associating the target with the original image in the optimal reduction ratio image; The method of obtaining the original image in the large scene and setting the reduction ratio level according to the size of the original image includes the following steps: Obtaining the original image size, wherein the original image size includes the height and width of the original image; Dividing the size lower than the original image size based on a preset size interval to obtain multiple size levels; Dividing the original image size by a plurality of size levels to obtain a plurality of reduction ratio levels corresponding to the plurality of size levels; The step of extracting target features from the original image after being processed at each reduction scale level, wherein the target features include the number of targets, the size of targets, and the overlap rate of target bounding boxes, comprises the following steps: The original image is reduced according to each reduction ratio level to obtain a reduced image corresponding to each reduction ratio level; At each reduction ratio level, using a target detection algorithm to extract target features in each reduced image, wherein the target features include the number of targets contained in each reduced image, the target size, and the overlap rate of the target bounding box; The comprehensive comparison of target features at each reduction ratio level and determination of the best reduction ratio according to the comparison result comprises the following steps: Obtaining current application requirements, and sorting the number of targets, target sizes, and overlapping rates of target bounding boxes based on the current application requirements to obtain a comparison order; Based on a preset threshold, and in accordance with the comparison order, the number of objects, the size of objects, and the overlap rate of the object bounding boxes at each reduction ratio level are compared in sequence, wherein the preset threshold includes a number threshold, a size threshold, and an overlap rate threshold corresponding to the number of objects, the size of objects, and the overlap rate of the object bounding boxes, respectively; Obtaining at least one pair of reduction ratio levels in which the number of objects, the size of objects, and the overlap rate of the object bounding boxes can respectively satisfy a number threshold, a size threshold, and an overlap rate threshold; Selecting a pair of highest reduction ratio levels from at least one pair of reduction ratio levels, wherein the higher the reduction ratio level, the greater the reduction ratio, and the less information the corresponding reduced image contains; The higher reduction ratio level of a pair of highest-ranked reduction ratio levels is selected as the optimal reduction ratio.

2. The large scene target processing method based on image features according to claim 1 is characterized in that: The current application requirement is a purpose of target detection, wherein the purpose of target detection includes counting requirements, location requirements and measurement requirements.

3. The large scene target processing method based on image features according to claim 2 is characterized in that: If the current application requirement is a counting requirement, the comparison order is the number of targets, target size, and the overlap rate of target bounding boxes; If the current application requirement is a location requirement, the comparison order is the overlap rate of the target bounding box, the target size, and the number of targets; If the current application requirement is a measurement requirement, the comparison order is target size, target bounding box overlap rate, and target quantity.

4. The large scene target processing method based on image features according to claim 3 is characterized in that: The method of reducing the original image according to the optimal reduction ratio to obtain the optimal reduction ratio image, and associating the target with the original image in the optimal reduction ratio image, comprises: Reducing the original image according to an optimal reduction ratio to obtain an image to be processed, wherein the image to be processed includes size information; Based on the size information of the image to be processed, the image to be processed is input into a corresponding target area search model, and at least one target area is determined according to a generation result of the target area search model, wherein the target area search model is a neural network model trained by a set training method, and the input for training the target area search model is image training samples at different size levels; Input each target area into the target identifier for position detection, and determine the target position frame in each target area; According to the size information and the best reduction ratio of each target area, the corresponding target position box is associated with the original image.

5. A large scene target processing system based on image features, applied to the target processing method according to any one of claims 1 to 4, characterized in that: The processing system includes: A reduction ratio level determination module is used to obtain the original image in a large scene and set the reduction ratio level according to the size of the original image; A target feature extraction module, used to extract target features in the original image after being processed at each reduction scale level, wherein the target features include the number of targets, the size of targets and the overlap rate of target bounding boxes; A comprehensive comparison module is used to comprehensively compare the target features at each reduction ratio level and determine the best reduction ratio based on the comparison results; The target association module is used to reduce the original image according to the optimal reduction ratio to obtain the optimal reduction ratio image, and associate the target with the original image in the optimal reduction ratio image.