An object-oriented changing target detection method and system based on improved YOLO

Abstract

The present application discloses an object-oriented change target detection method and system based on improved YOLO; the method comprises: collecting image data of target detection areas at different times, aligning image data at two different times, generating image pairs of target detection areas, inputting the pre-processed data into an improved YOLOV8 model, utilizing two input branches of a shared backbone to extract features of interest from an input image respectively, utilizing the DCAM attention module in the improved YOLOV8 model to enhance and fuse the feature maps of interest extracted by the two branches, and obtaining a difference feature map between the fused features and the original features; utilizing the neck in the improved YOLOV8 model to perform multi-scale fusion processing on the obtained difference feature map; utilizing the detection head in the improved YOLOV8 model to identify and locate targets with differences, and outputting images with the difference targets of interest between the image pairs annotated in the form of bounding boxes; the present application can achieve high-precision detection of object changes while improving the speed and robustness of change detection.

Description

Technical Field

[0001] The present application relates to the technical field of target detection, and specifically to an object-oriented change target detection method and system based on improved YOLO. Background Art

[0002] Change detection technology has been widely used in remote sensing image analysis, particularly in areas such as construction progress monitoring and urban planning and development. Traditional change detection methods rely on manual feature extraction and complex image processing algorithms. However, with the recent development of deep learning technology, particularly the maturity of object detection models, change detection technology has gradually moved towards automation and intelligence. These new methods not only improve detection accuracy but also significantly reduce the need for manual intervention, thereby increasing efficiency and accuracy.

[0003] Among existing change detection technologies, commonly used methods include pixel-level change detection, object-level change detection, and scene-level change detection. Pixel-level change detection generates a difference map by comparing each pixel in each pair of images, and is suitable for applications that require detailed change information. Object-level change detection focuses on changes in specific semantic objects, such as the addition or removal of buildings, and usually uses bounding boxes or masks to mark the changed targets. Scene-level change detection pays more attention to changes in the overall environment, such as urban expansion or land use changes. In order to improve detection accuracy and efficiency, researchers usually adopt a combination of multiple technical means, such as image registration, feature extraction, attention mechanism, etc.

[0004] However, existing change detection methods still have some shortcomings in practical applications. For example, traditional pixel-level methods are easily affected by factors such as lighting and shadows, resulting in high false detection and missed detection rates. Although object-level methods can better capture specific changing targets, in complex backgrounds, especially in changing natural environments, the accuracy of object-level methods in feature extraction and target recognition still needs to be improved. In addition, existing methods consume a lot of computing resources when processing large-scale data sets, making it difficult to achieve real-time or near-real-time detection requirements. Therefore, how to achieve high-precision detection of object changes while improving the speed and robustness of change detection has become a key issue that needs to be addressed urgently. Summary of the Invention

[0005] In order to achieve high-precision detection of object changes while improving the speed and robustness of change detection, the present application provides an object-oriented change target detection method and system based on improved YOLO.

[0006] In the first aspect, an object-oriented change target detection method based on improved YOLO includes:

[0007] Collect image data of the target detection area at different times, align the image data at two different times, and generate an image pair of the target detection area;

[0008] Preprocessing the image pairs of the target detection area;

[0009] The preprocessed image pair of the target detection area is input into the improved YOLOV8 model, and the two input branches sharing the backbone in the improved YOLOV8 model are used to extract the features of interest from one image in the input image pair respectively. The feature maps of interest extracted by the two branches are enhanced and fused using the DCAM attention module in the improved YOLOV8 model to obtain the difference feature map between the fused features and the original features; the neck in the improved YOLOV8 model is used to perform multi-scale fusion processing on the obtained difference feature map; the detection head in the improved YOLOV8 model is used to identify and locate targets with differences, and output images with bounding boxes marking the different targets of interest between the image pairs; the structural improvements of the improved YOLO model include: the output of the model is designed to be two branches, each branch shares the backbone, and the DCAM attention module is introduced in the neck part of the YOLO V8 model.

[0010] By adopting the above scheme, based on the traditional YOLOV8 model, a dual-branch input is designed to extract features from the registered images. The shared backbone design reduces redundant calculations and improves model efficiency. The DCAM attention module is introduced to enhance the expressiveness of feature maps, enabling the model to better capture the key features of target changes. Multi-scale fusion processing further enhances the model's ability to detect targets of different scales, effectively improving the accuracy and robustness of object change detection and adapting to change detection needs in various complex scenarios.

[0011] Preferably, the structural improvement of the improved YOLO V8 model further includes: introducing a pre-processing module ShuffleNet into the YOLO V8 model structure, using ShuffleNet to perform scene classification on the input image of the target detection area to obtain a scene classification label; the scene classification includes office scenes and outdoor scenes;

[0012] The process of enhancing and fusing the feature maps of interest extracted by the two branches using the DCAM attention module in the improved YOLOV8 model also includes: querying the matching attention weight modulation parameters from the first scene prior knowledge base according to the acquired scene classification label, modulating the attention weight of the DCAM according to the queried attention weight modulation parameters, and enhancing and fusing the feature maps of interest extracted by the two branches according to the modulated attention weight; the first scene prior knowledge base stores the preset attention weight adjustment parameters corresponding to each scene, and the preset attention weight adjustment parameters are determined according to the attention adjustment parameters corresponding to the optimal target detection feedback results under the conditions of different preset attention weight adjustment parameters for different scenes.

[0013] By adopting the above scheme, a lightweight network is set as the front module to obtain image scene classification, and the DCAM attention weight is modulated according to the scene classification label, so that the model can better adapt to the change detection requirements in specific scenarios when processing different scenes, thereby improving the detection performance.

[0014] Preferably, the multi-scale fusion processing of the acquired difference feature map using the neck in the improved YOLOV8 model further includes:

[0015] According to the obtained scene classification label, the matching multi-scale fusion weight modulation parameters are queried from the second scene prior knowledge base, the multi-scale fusion weight parameters in the neck are adjusted according to the queried multi-scale fusion weight modulation parameters, and the difference feature map is multi-scale fused according to the modulated multi-scale fusion weight; the second scene prior knowledge base stores the preset multi-scale fusion parameters corresponding to each scene, and the preset multi-scale fusion parameters are determined according to the multi-scale fusion parameters corresponding to the optimal target detection feedback results under the conditions of different preset multi-scale fusion parameters for different scenes.

[0016] By adopting the above scheme, matching the weight parameters of multi-scale fusion according to different scenes and adjusting the weight parameters of multi-scale fusion in the neck, we can better capture and fuse features of different scales, reduce detection errors caused by scene changes, and improve the overall performance of target detection.

[0017] Preferably, the step of using the detection head in the improved YOLOV8 model to identify and locate targets with differences, and outputting images with bounding boxes marking the targets of interest between the image pairs further includes:

[0018] According to the obtained scene classification label, the modulation parameter is adjusted by querying the matching confidence threshold from the third scene prior knowledge base, the confidence threshold in the detection head is adjusted according to the queried confidence modulation parameter, and the difference target detection is performed on the difference feature map according to the modulated confidence threshold; the third scene prior knowledge base stores the preset confidence threshold corresponding to each scene, and the preset confidence threshold is determined according to the confidence threshold corresponding to the optimal target detection feedback result under different confidence threshold conditions in different historical scenes.

[0019] By adopting the above solution, the parameters are adjusted according to the confidence threshold of the scene classification label query match, the confidence threshold of the detection head is dynamically adjusted, the performance of the detection head is optimized, false detections and missed detections are reduced, and the detection effect is improved.

[0020] Preferably, the preprocessing includes: analyzing the light intensity and distribution of the images of the target detection area at different times, using image processing technology to perform light normalization processing on the images at two times, and making the light intensity and distribution of the images of the target detection area at different times after the light normalization processing consistent.

[0021] By adopting the above scheme, the image is subjected to light normalization processing so that the light intensity and distribution of the processed image are consistent, thus eliminating the pseudo changes caused by light changes and improving the accuracy and robustness of change detection.

[0022] Preferably, it also includes:

[0023] Before outputting an image with a target of interest between a pair of images marked in the form of a bounding box, image light information is determined for the image with the target of interest between the pair of images marked in the form of a bounding box, and original light information is retrieved from a light mapping table. Light of the image with the target of interest between the pair of images marked in the form of a bounding box is restored based on the retrieved original light information; the light mapping table stores the original light information of the original image before light normalization processing and pre-processed light information mapped and associated with the original light information, and the light information includes: light intensity and distribution.

[0024] By adopting the above scheme, the image that has been subjected to light normalization processing is restored in the output stage to ensure that the output image retains the true visual effect under the original lighting conditions, thereby improving the accuracy and reliability of the change detection results.

[0025] Preferably, the shared backbone structure includes: 5 convolutional layers and 4 pooling layers.

[0026] By adopting the above solution, the shared backbone is simplified and the model's efficiency in detecting target changes is improved.

[0027] In a second aspect, the present application provides an object-oriented change target detection system based on improved YOLO, comprising: an image pair acquisition module for collecting image data of a target detection area at different times, registering the image data at two different times, and generating an image pair of the target detection area;

[0028] An image pair preprocessing module, used for preprocessing the image pairs of the target detection area;

[0029] An image-to-object change detection module is configured to input the preprocessed image pair into the improved YOLOV8 model, extract features of interest from one image in the input image pair using the two input branches that share the backbone in the improved YOLOV8 model, enhance and fuse the feature maps of interest extracted by the two branches using the DCAM attention module in the improved YOLOV8 model, and obtain a difference feature map between the fused features and the original features; perform multi-scale fusion processing on the obtained difference feature map using the neck in the improved YOLOV8 model; identify and locate targets with differences using the detection head in the improved YOLOV8 model, and output an image with bounding boxes marking the targets of interest between the image pairs; structural improvements to the improved YOLO model include: designing the model output as two branches, each branch sharing the backbone, and introducing the DCAM attention module in the neck part of the YOLOV8 model.

[0030] By adopting the above scheme, a dual-branch structure is designed to extract features from image pairs. The shared backbone and DCAM attention modules in the improved YOLOV8 model are utilized to enhance the feature extraction and fusion capabilities, improve the accuracy and robustness of detection, and achieve efficient object-oriented change target detection by outputting the difference targets of interest between image pairs in the form of bounding boxes through the detection head.

[0031] In a third aspect, the present application provides a computer-readable storage medium, which includes a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute the method as described above.

[0032] In a fourth aspect, the present application provides a computer device, which includes a memory, a processor, and a program stored and executable on the memory, and the program implements the steps of the above method when executed by the processor.

[0033] In summary, this application has the following beneficial effects:

[0034] 1. Based on the YOLOV8 model, a dual-branch structure is designed with a shared backbone to extract features from input image pairs. Combined with the introduced DCAM attention module, it can effectively extract and fuse features from the two images, significantly improving the accuracy and robustness of target change detection.

[0035] 2. The pre-processing module ShuffleNet is introduced for scene classification. It combines multiple scene prior knowledge bases to dynamically adjust attention weights, multi-scale fusion weights, and confidence thresholds, achieving adaptive optimization for different scenarios and further improving detection accuracy and reliability.

[0036] 3. Through light normalization processing and light restoration operations, the problems of false detection and missed detection caused by light changes are eliminated, the consistency and stability of the detection results are ensured, and the practicality and reliability of the entire system are improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 Flowchart of the object-oriented change target detection method based on improved YOLO described in a specific embodiment;

[0038] Figure 2 Schematic diagram of a 6-channel input model structure in the object-oriented changing target detection method based on improved YOLO described in a specific embodiment;

[0039] Figure 3 Schematic diagram of introducing the DCMA module into the object-oriented change target detection method based on improved YOLO described in a specific embodiment;

[0040] Figure 4 Schematic diagram of the results obtained by using the object-oriented change target detection method based on the improved YOLO described in the specific embodiment; Figure 4 (a) Schematic diagram of indoor results obtained using the object-oriented change target detection method based on the improved YOLO described in the specific embodiment; Figure 4 (b) Schematic diagram of outdoor results obtained using the object-oriented change target detection method based on the improved YOLO described in the specific embodiment;

[0041] Figure 5 Schematic diagram of the structure of the object-oriented change target detection system based on the improved YOLO described in the specific embodiment. DETAILED DESCRIPTION

[0042] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0043] like Figure 1 As shown, the embodiment of the present application discloses an object-oriented change target detection method based on improved YOLO, including:

[0044] S1. Collect image data of the target detection area at different times to generate image pairs of the target detection area.

[0045] Specifically, the image data of the target detection area is collected by using a camera device, and the collected image data of the target area can be framed and processed, and the image data of every two consecutive frames can be aligned in real time according to the time sequence to generate several image pairs of the target monitoring area; or according to user needs, the image data frames of the target detection area at any two moments of pre-detection of the target detection area can be used to generate image pairs of the target monitoring area.

[0046] The registration of image data at two different moments can be performed using a feature-based registration method, a grayscale-based registration method, a transform domain-based registration method, or a deep learning-based registration method. In this embodiment, a feature matching algorithm, such as SIFT or SURF, is used to extract key points from the image and perform matching. False matching points are then removed using the RANSAC algorithm, and finally, the images are aligned using an affine transformation or a homography transformation.

[0047] S2. Preprocess the image of the target detection area.

[0048] Specifically, the preprocessing includes image denoising, image enhancement, and size normalization.

[0049] S3. Input the preprocessed pairs into the improved YOLOV8 model, and output an image with the difference of interest between the image pairs marked in the form of bounding boxes.

[0050] Specifically, the improved YOLOV8 model is based on the YOLOV8 model with some structural improvements. The main structural improvements include:

[0051] The input of the model adopts the 6-channel output structure commonly used in existing technologies, such as Figure 2As shown in the figure, a feature extraction operation is performed on the input image pair to obtain the feature maps corresponding to the two images in the image pair respectively; on this basis, the input structure is designed as a dual-branch structure, and the two branch structures share a backbone. The backbone structure includes 5 convolutional layers and 4 convolutional layers. Batch normalization (Batch Normalization) and activation function (ReLU) are added between each convolutional layer and pooling layer of the backbone to improve the convergence speed and generalization ability of the model. A convolution operation is performed on one image data in the input image pair with a convolutional structure of 1 CNN+BN+relu and 4 (CNN+BN+relu)×2; 4 pooling layers are interspersed in the 5 convolutional layers, and pooling operations are performed on the feature data extracted by each convolutional layer.

[0052] The DCAM attention module is introduced into the neck part of the YOLO V8 model. The DCAM attention module includes a spatial attention mechanism and a channel attention mechanism. The DCAM attention module is used to learn and assign attention weights. By weighting the specific spatial positions in the feature map of each branch, the most critical areas for the task are highlighted to achieve feature enhancement. The features in the feature maps extracted by the two branches are then weighted and fused to obtain a fused feature map. The fused feature map is compared with the two original feature maps to determine the difference features whose feature differences are greater than the preset differences, and generate a difference feature map. The specific DCAM-assigned attention weight distribution is determined through model training.

[0053] The improved YOLOV8 model training can be performed by training historical image pairs that are annotated with the classification and location of the difference targets of interest to the user; the input of the model is an image pair, and the output is an image that is annotated with the difference targets of interest between the image pairs in the form of bounding boxes, where the bounding boxes are located in the image that is later in time order.

[0054] like Figure 3As shown, the process of obtaining the output annotation of the difference target of interest using the improved YOLOV8 model includes: using the two input branches of the shared backbone to extract the features of interest from one image in the input image pair respectively, using the DCAM attention module in the improved YOLOV8 model to enhance and fuse the feature maps of interest extracted by the two branches, and obtain the difference feature map between the fused features and the original features. Specifically, the feature maps of interest of the same scale at multiple different scales extracted by the two branches are enhanced and fused to generate difference feature maps of different scales; using the neck in the improved YOLOV8 model to perform multi-scale fusion processing on the obtained difference feature maps; using the detection head in the improved YOLOV8 model to identify and locate targets with differences. Specifically, the detection head can adopt the anchor box mechanism to preliminarily screen the candidate targets through the pre-set anchor box, and then remove the overlapping candidate boxes through the non-maximum suppression (NMS) algorithm, and output the image with the bounding box annotated with the difference target of interest between the image pairs.

[0055] Using the method described in the above embodiment, as Figure 4 As shown in the figure, the improved YOLOV8 model can be used to detect changes in multiple categories of objects in image pairs. While maintaining high-precision detection, the speed and robustness of object change detection are improved. The test results are shown in Figure 4 (a) with Figure 4 (b).

[0056] In a specific embodiment, in order to further improve the speed and robustness of object change detection and adapt to object change detection in complex and changing scenes, the method further includes:

[0057] Taking into account the different importance of the same target in different scenarios, in a specific or typical scenario, the importance of some targets needs to be adjusted accordingly. In order to adapt to the changing scenarios, the parameters of these targets can be adjusted in advance to improve the detection accuracy and robustness of the model.

[0058] Specifically, a pre-processing module ShuffleNet is introduced into the YOLO V8 model structure. ShuffleNet is used to perform scene classification on the input image of the target detection area to obtain scene classification labels, thereby providing effective prior information for subsequent feature extraction and target detection. In this embodiment, the scene classification includes office scenes and outdoor scenes.

[0059] The process of enhancing and fusing the feature maps of interest extracted by the two branches using the DCAM attention module in the improved YOLOV8 model also includes:

[0060] The matching attention weight modulation parameters are queried from the first scene prior knowledge base based on the acquired scene classification labels. The first scene prior knowledge base stores the preset attention weight adjustment parameters corresponding to each scene, such as: the office scene pays more attention to the changes of indoor objects, while in the outdoor scene, it may pay more attention to the changes in terrain and vegetation, and the optimal attention weight ratio of the changing target (specific target) that is more concerned is set accordingly, and the set attention weight ratio is preset as the preset attention weight adjustment parameter, such as: the preset first attention weight adjustment parameter corresponding to the office scene, and the preset second attention weight adjustment parameter corresponding to the outdoor scene; the preset attention weight adjustment parameter is determined according to the attention adjustment parameter corresponding to the optimal target detection feedback result under the conditions of different preset attention weight adjustment parameters for different scenes. The optimal target detection feedback result refers to the target detection feedback result with the greatest similarity with the actual result during the target detection process of the user simulation test.

[0061] The attention weight of DCAM is modulated according to the queried attention weight modulation parameter; the feature maps of interest extracted by the two branches are enhanced and fused according to the modulated attention weight;

[0062] By using the method described in the above embodiment, the attention weight can be dynamically adjusted in different scenarios to improve the detection accuracy and robustness of the model.

[0063] In a specific embodiment, in order to further improve the speed and robustness of object change detection and adapt to the requirements of object change detection in complex and changing scenes, the method further includes:

[0064] The multi-scale fusion processing of the obtained difference feature map using the neck in the improved YOLOV8 model also includes: querying the matching multi-scale fusion weight modulation parameters from the second scene prior knowledge base according to the obtained scene classification label. Wherein, the second scene prior knowledge base stores the preset multi-scale fusion parameters corresponding to each scene, such as: more small-scale features may be required, while in outdoor scenes, more large-scale features may be required, and the feature weight ratios of the fusion of features of different scales are set accordingly, and the feature weight ratios of the fusion of features of different scales are set as the preset multi-scale fusion weight modulation parameters, such as: the preset first multi-scale fusion weight modulation parameters corresponding to the office scene, and the preset second multi-scale fusion weight modulation parameters corresponding to the outdoor scene; the preset multi-scale fusion parameters are determined according to the multi-scale fusion parameters corresponding to the optimal target detection feedback results under the conditions of different preset multi-scale fusion parameters for different scenes.

[0065] The multi-scale fusion weight parameters in the neck are adjusted according to the queried multi-scale fusion weight modulation parameters, and the multi-scale fusion of the difference feature map is performed according to the modulated multi-scale fusion weights.

[0066] By using the method described in the above embodiment, the multi-scale fusion weights can be dynamically adjusted in different scenarios to improve the detection accuracy and robustness of the model.

[0067] In a specific embodiment, in order to improve the accuracy and reliability of detection, the method further includes:

[0068] The method of using the detection head in the improved YOLOV8 model to identify and locate targets with differences, and outputting images with bounding boxes marking the targets of interest between the image pairs, further includes:

[0069] The modulation parameters are adjusted based on the confidence threshold of the match queried from the third scene prior knowledge base according to the acquired scene classification label. The third scene prior knowledge base stores the preset confidence threshold for each scene match. For example, an outdoor scene may require a lower threshold due to the complex background, while an office scene may be suitable for a higher threshold due to the relatively stable environment. Correspondingly, a first preset confidence threshold corresponding to the outdoor scene and a second preset confidence threshold corresponding to the office scene are set, and the first confidence threshold is smaller than the second confidence threshold. Specifically, the preset confidence threshold is determined based on the confidence threshold corresponding to the optimal target detection feedback result under different confidence threshold conditions in different historical scenes.

[0070] The confidence threshold in the detection head is adjusted according to the queried confidence modulation parameter, and difference target detection is performed on the difference feature map according to the modulated confidence threshold.

[0071] By using the method described in the above embodiment, the confidence threshold can be dynamically adjusted in different scenarios, thereby improving the accuracy and reliability of model detection.

[0072] In a specific embodiment, considering the influence of light on image feature extraction and detection, in order to avoid the influence of light and improve the reliability and accuracy of the detection results, the method further includes:

[0073] The preprocessing includes: analyzing the light intensity and distribution of the images of the target detection area at different times, which can be specifically achieved by calculating the brightness histogram, contrast and other parameters of the image; using image processing techniques, such as histogram equalization and contrast stretching, to perform light normalization processing on the images at two times, and making the light intensity and distribution of the images of the target detection area at different times after the light normalization processing consistent.

[0074] In the preprocessing process, the original light information before the original image is subjected to light normalization processing and the preprocessed light information mapped thereto are stored as a pair of mapping relationships in a light mapping table, wherein the light information includes light intensity and distribution.

[0075] Furthermore, to enhance the visual quality of the detection results, a light mapping table is used to restore the light rays of the remote sensing images in which the difference targets of interest between the remote sensing image pairs are annotated with bounding boxes. Specifically, the method further comprises: before outputting the images in which the difference targets of interest between the image pairs are annotated with bounding boxes, using image analysis techniques to determine the light rays of the images in which the difference targets of interest between the image pairs are annotated with bounding boxes, querying the light mapping table for the original light rays, and restoring the light rays of the images in which the difference targets of interest between the image pairs are annotated with bounding boxes based on the retrieved original light rays.

[0076] like Figure 5 As shown, an object-oriented change target detection system based on improved YOLO includes:

[0077] An image pair acquisition module 101 is used to collect image data of a target detection area at different times, register the image data at two different times, and generate an image pair of the target detection area;

[0078] An image pair preprocessing module 102 is used to preprocess the image pair of the target detection area;

[0079] The image-to-object change detection module 103 is used to input the preprocessed image pair of the target detection area into the improved YOLOV8 model, use the two input branches of the improved YOLOV8 model that share the backbone to extract features of interest from one image in the input image pair respectively, use the DCAM attention module in the improved YOLOV8 model to enhance and fuse the feature maps of interest extracted by the two branches to obtain a difference feature map between the fused features and the original features; use the neck in the improved YOLOV8 model to perform multi-scale fusion processing on the obtained difference feature map; use the detection head in the improved YOLOV8 model to identify and locate targets with differences, and output images with the difference targets of interest between the image pairs marked in the form of bounding boxes; wherein, the structural improvements of the improved YOLO model include: the output of the model is designed to be two branches, each branch shares the backbone, and the DCAM attention module is introduced in the neck part of the YOLO V8 model.

[0080] The embodiment of the present application also discloses a computer-readable storage medium.

[0081] Specifically, the computer-readable storage medium stores a computer program that can be loaded by a processor and executed as described above, such as the object-oriented change target detection method based on the improved YOLO. The computer-readable storage medium includes, for example, various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0082] The embodiment of the present application also discloses a computer device.

[0083] Specifically, the computer device includes a memory and a processor, and the memory stores a computer program that can be loaded by the processor and execute the above-mentioned object-oriented change target detection method based on the improved YOLO.

[0084] The above are all preferred embodiments of the present application and are not intended to limit the scope of protection of this application. Unless otherwise stated, any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features. In other words, unless otherwise stated, each feature is merely an example of a series of equivalent or similar features.

Claims

1. An object-oriented change target detection method based on improved YOLO, characterized in that: include: Collect image data of the target detection area at different times, align the image data at two different times, and generate an image pair of the target detection area; Preprocessing the image pairs of the target detection area; The preprocessed image pair of the target detection area is input into the improved YOLOV8 model. The two input branches of the shared backbone in the improved YOLOV8 model are used to extract the features of interest from one image in the input image pair respectively. The DCAM attention module in the improved YOLOV8 model is used to enhance and fuse the feature maps of interest extracted by the two branches to obtain the difference feature map between the fused features and the original features. The neck in the improved YOLOV8 model is used to perform multi-scale fusion processing on the acquired difference feature maps; The improved YOLO V8 model uses a detection head to identify and locate objects with differences, and outputs images with bounding boxes annotating the objects of interest between the image pairs. The improved YOLO model includes designing the model output into two branches, each branch sharing a backbone, and introducing a DCAM attention module into the neck portion of the YOLO V8 model. The improved YOLO V8 model also includes introducing a pre-processing module, ShuffleNet, into the YOLO V8 model structure, and using ShuffleNet to perform scene classification on the input image of the target detection area to obtain a scene classification label. The scene classification includes office scenes and outdoor scenes. The process of enhancing and fusing the feature maps of interest extracted by the two branches using the DCAM attention module in the improved YOLOV8 model also includes: querying a matching attention weight modulation parameter from a first scene prior knowledge base according to the acquired scene classification label, modulating the attention weight of the DCAM according to the queried attention weight modulation parameter, and enhancing and fusing the feature maps of interest extracted by the two branches according to the modulated attention weight; the first scene prior knowledge base stores preset attention weight adjustment parameters corresponding to each scene, and the preset attention weight adjustment parameters are determined according to the attention adjustment parameters corresponding to the optimal target detection feedback results under different preset attention weight adjustment parameters for different scenes; The method of using the detection head in the improved YOLOV8 model to identify and locate targets with differences, and outputting images with bounding boxes marking the targets of interest between the image pairs, further includes: According to the obtained scene classification label, the modulation parameter is adjusted by querying the matching confidence threshold from the third scene prior knowledge base, the confidence threshold in the detection head is adjusted according to the queried confidence modulation parameter, and the difference target detection is performed on the difference feature map according to the modulated confidence threshold; the third scene prior knowledge base stores the preset confidence threshold corresponding to each scene, and the preset confidence threshold is determined according to the confidence threshold corresponding to the optimal target detection feedback result under different confidence threshold conditions in different historical scenes.

2. The object-oriented change target detection method based on improved YOLO according to claim 1, characterized in that: The multi-scale fusion processing of the obtained difference feature map using the neck in the improved YOLOV8 model also includes: According to the obtained scene classification label, the matching multi-scale fusion weight modulation parameters are queried from the second scene prior knowledge base, the multi-scale fusion weight parameters in the neck are adjusted according to the queried multi-scale fusion weight modulation parameters, and the difference feature map is multi-scale fused according to the modulated multi-scale fusion weight; the second scene prior knowledge base stores the preset multi-scale fusion parameters corresponding to each scene, and the preset multi-scale fusion parameters are determined according to the multi-scale fusion parameters corresponding to the optimal target detection feedback results under the conditions of different preset multi-scale fusion parameters for different scenes.

3. The object-oriented change target detection method based on improved YOLO according to claim 1, characterized in that: The preprocessing includes: analyzing the light intensity and distribution of the images of the target detection area at different times, using image processing technology to perform light normalization processing on the images at two times, and making the light intensity and distribution of the images of the target detection area at different times after the light normalization processing consistent.

4. The object-oriented change target detection method based on improved YOLO according to claim 3 is characterized in that: Also includes: Before outputting the image with the target of interest between the image pair being annotated in the form of a bounding box, determining image light information for the image with the target of interest between the image pair being annotated in the form of a bounding box, querying the original light information from a light mapping table, and restoring the light of the image with the target of interest between the image pair being annotated in the form of a bounding box based on the query original light information; The light mapping table stores original light information of the original image before light normalization processing and pre-processed light information associated with the original image, and the light information includes light intensity and distribution.

5. The object-oriented change target detection method based on improved YOLO according to claim 1, characterized in that: The shared backbone structure includes 5 convolutional layers and 4 pooling layers.

6. An object-oriented change target detection system based on improved YOLO, the system being used to implement the method according to any one of claims 1 to 5, characterized in that: include: An image pair acquisition module is used to collect image data of the target detection area at different times, align the image data at two different times, and generate an image pair of the target detection area; An image pair preprocessing module, used for preprocessing the image pairs of the target detection area; The image-to-object change detection module is used to input the preprocessed image pairs of the target detection area into the improved YOLOV8 model, use the two input branches of the shared backbone in the improved YOLOV8 model to extract the features of interest from one image in the input image pair, and use the DCAM attention module in the improved YOLOV8 model to enhance and fuse the feature maps of interest extracted by the two branches to obtain the difference feature map between the fused features and the original features; The neck in the improved YOLOV8 model is used to perform multi-scale fusion processing on the acquired difference feature maps; The detection head in the improved YOLOV8 model is used to identify and locate objects with differences, and output images with bounding boxes annotated with the objects of interest between the image pairs. The structural improvements of the improved YOLO model include: the output of the model is designed to be two branches, each branch shares the backbone, and the DCAM attention module is introduced in the neck part of the YOLO V8 model.

7. A computer-readable storage medium, characterized in that The computer-readable storage medium includes a stored computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located is controlled to execute the method according to any one of claims 1 to 5.

8. A computer device, characterized in that: The computer device includes a memory, a processor, and a program stored and executable on the memory, and when the program is executed by the processor, the steps of the method according to any one of claims 1 to 5 are implemented.

Images