Target detection method, apparatus, device, and medium

Abstract

The application relates to the technical field of image processing, and discloses a target detection method, a target detection device, equipment and a medium. The method comprises the following steps: acquiring a panoramic image to be detected; based on key image features in the panoramic image to be detected, adaptively slicing the panoramic image to be detected to obtain a plurality of original slice images; supplementing neighborhood information of the original slice images to obtain a plurality of supplemented slice images; performing target detection on the supplemented slice images to generate corresponding target detection boxes; and performing global consistency checking on the target detection boxes to eliminate target detection boxes that do not meet preset consistency conditions, thereby obtaining a target detection result corresponding to the panoramic image to be detected. The embodiment of the application can adapt to various panoramic image slicing scenes and improve the quality of panoramic image slicing.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to a target detection method, apparatus, device, and medium. Background Technology

[0002] When performing object detection on panoramic images, directly performing object detection on the panoramic image not only consumes a large amount of computing resources and leads to low detection efficiency, but also easily results in insufficient target feature extraction due to the mismatch between image resolution and the input size of the detection model. Therefore, related technologies generally adopt a technical approach of slicing combined with local detection. That is, the panoramic image is first sliced ​​into multiple undistorted local planar images, then object detection is performed on each local planar image, and finally the detection results are summarized. Existing panoramic image slicing techniques are mainly divided into fixed-size slicing and simple regularized adaptive slicing. However, existing panoramic image slicing techniques rely on manual experience to set the slice size or slicing rules, which can only cover a small number of known scenes and cannot adapt to complex panoramic images with complex content and random target distribution. Summary of the Invention

[0003] The purpose of this application is to provide a target detection method, apparatus, device, and medium that can adapt to various panoramic image tiling scenarios and improve the quality of panoramic image tiling through a mechanism of dynamic adaptive tiling and associated tiling information.

[0004] This application provides a target detection method, including:

[0005] Acquire the panoramic image to be detected;

[0006] Based on the key image features in the panoramic image to be detected, adaptive slicing is performed on the panoramic image to be detected to obtain several original slice images.

[0007] The original slice image is supplemented with neighborhood information to obtain several supplementary slice images;

[0008] Target detection is performed on the supplementary slice image to generate corresponding target detection boxes;

[0009] A global consistency check is performed on the target detection boxes to remove target detection boxes that do not meet the preset consistency conditions, thereby obtaining the target detection results corresponding to the panoramic image to be detected.

[0010] In some embodiments, the adaptive slicing of the panoramic image to be detected based on key image features in the panoramic image to be detected includes:

[0011] Key features are extracted from the panoramic image to be detected to obtain the corresponding target of interest features, local region features, and global image features;

[0012] Based on the mapping relationship between the key image features and the slicing parameters, adaptive slicing parameters corresponding to the target of interest features, the local region features, and the global image features are determined;

[0013] Based on the adaptive slicing parameters, several preliminary slice images are generated;

[0014] The initial slice image is deduplicated and fine-tuned to obtain the original slice image.

[0015] In some embodiments, the deduplication and fine-tuning of the preliminary slice image includes:

[0016] Based on the size characteristics of the target to be detected, the overlap rate between the initial slice images is configured;

[0017] When the edge of the preliminary slice image passes through the target to be detected, the position of the preliminary slice image is finely adjusted so that the preliminary slice image contains the complete target to be detected;

[0018] Remove the initial slice image containing only invalid regions.

[0019] In some embodiments, supplementing the original slice image with neighborhood information includes:

[0020] Extract the target contour features of the target to be detected from the original slice image;

[0021] Calculate the contour similarity between the target contour features of adjacent original slice images;

[0022] When the contour similarity meets the preset contour similarity conditions, feature mapping is performed on the target contour features so that the non-overlapping target contour features of the target to be detected are supplemented between adjacent original slice images to obtain the supplementary slice image.

[0023] In some embodiments, supplementing the original slice image with neighborhood information includes:

[0024] Extract overlapping and non-overlapping texture features between adjacent original slice images;

[0025] Calculate the overlapping texture similarity between overlapping texture features of adjacent original slice images;

[0026] When the overlapping texture similarity meets the preset texture similarity condition, feature mapping is performed on the non-overlapping texture features so that the non-overlapping texture features are mutually supplemented between adjacent original slice images to obtain the supplementary slice image.

[0027] In some embodiments, performing a global consistency check on the target detection box includes:

[0028] Obtain the target category information of the target to be detected;

[0029] Based on the target category information, the target detection boxes are subjected to size and shape verification. Target detection boxes whose size verification results do not meet the preset size verification conditions and whose shape verification results do not meet the preset shape verification conditions are removed.

[0030] In some embodiments, performing a global consistency check on the target detection box includes:

[0031] Acquire the target location information and target spatial information of the target to be detected;

[0032] Based on the target location information and the target spatial information, the target detection box is positionally verified, and target detection boxes that deviate from the position represented by the target location information and the region represented by the target spatial information are removed.

[0033] This application also provides a target detection device, including:

[0034] The first module is used to acquire the panoramic image to be detected;

[0035] The second module is used to adaptively slice the panoramic image to be detected based on the key image features in the panoramic image to be detected, and obtain several original slice images.

[0036] The third module is used to supplement the original slice image with neighborhood information to obtain several supplementary slice images;

[0037] The fourth module is used to perform target detection on the supplementary slice image to generate corresponding target detection boxes;

[0038] The fifth module is used to perform global consistency verification on the target detection boxes to remove target detection boxes that do not meet the preset consistency conditions, thereby obtaining the target detection result corresponding to the panoramic image to be detected.

[0039] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described target detection method.

[0040] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described target detection method.

[0041] The beneficial effects of this application are as follows: By adapting content features to adaptive tiling based on key image features, it avoids target fragmentation and context loss caused by fixed tiling. By supplementing with neighborhood information to establish an inter-slice association mechanism, it effectively alleviates the false detection problem caused by the fragmentation of boundary information. Combined with global consistency verification to verify the rationality of detection results, it significantly improves the reliability of detection results. It has the advantages of improving the accuracy and efficiency of target detection, avoiding target fragmentation and context loss, reducing false detection and invalid tiling, and can adapt to various panoramic image tiling scenarios, thereby improving the quality of panoramic image tiling. Attached Figure Description

[0042] Figure 1 This is a diagram illustrating the application environment of the target detection method provided in the embodiments of this application.

[0043] Figure 2 This is a flowchart of the target detection method provided in the embodiments of this application.

[0044] Figure 3 This is a flowchart of a method for adaptive slicing of a panoramic image to be detected, provided in an embodiment of this application.

[0045] Figure 4 This is a flowchart of a method for supplementing neighborhood information of an original slice image according to the first embodiment of this application.

[0046] Figure 5 This is a flowchart of a method for supplementing neighborhood information of an original slice image according to the second embodiment of this application.

[0047] Figure 6 This is a flowchart of a method for performing global consistency verification of target detection boxes provided in the first embodiment of this application.

[0048] Figure 7 This is a flowchart of a method for performing global consistency verification of target detection boxes provided in the second embodiment of this application.

[0049] Figure 8 This is a schematic diagram of the target detection device provided in the embodiments of this application.

[0050] Figure 9 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0052] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and drawings are used to distinguish similar objects and are not used to describe a specific order or sequence.

[0053] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0054] The target detection method provided in this application can be executed by a computer device, which can be a terminal device or a server. The terminal device includes, but is not limited to, mobile phones, computers, smart home appliances, vehicle terminals, and aircraft. The server can be a standalone physical server, a server cluster consisting of multiple physical servers, a distributed system, or a cloud server. Furthermore, the information, data, and signals involved in this application's embodiments are all authorized by the relevant parties or have been fully authorized by all parties, and the collection, use, and processing of related data comply with the relevant laws, regulations, and standards of the relevant countries and regions.

[0055] In panoramic image target detection processes, such as those used in power line inspection, ultra-high resolution image processing faces challenges of excessive computational resource consumption and reduced detection efficiency. Specifically, image pixel sizes typically exceed 10000×5000, significantly increasing the computational load for direct target detection. Furthermore, the mismatch between image resolution and the input size of the detection model leads to insufficient target feature extraction. Existing slicing techniques lack dynamic adaptation to image content features, resulting in large targets being fragmented into multiple pieces while small targets suffer from missing contextual information due to excessively large slice sizes. In areas with uneven target distribution, fixed-size slicing strategies cannot adjust to differences in target density, leading to feature overlap in densely populated areas and redundant slices in sparsely populated areas, further reducing detection accuracy. Moreover, the lack of information fragmentation between slices and a global consistency verification mechanism prevents cross-slice correlation verification of local detection results, causing false detections due to contradictory location logic.

[0056] For example, in substation facility inspection scenarios, the acquired panoramic images contain small openings and defects with diameters of 15-30 pixels and large edge structures exceeding 100 pixels. When using a fixed-size slicing method, small openings lack sufficient contextual information due to insufficient slice coverage, making them difficult to identify accurately; large edges are segmented into multiple independent fragments, resulting in incomplete detection results. In densely populated transformer areas, targets within slices overlap, leading to false detections; in open corridor areas, redundant slices with numerous target-free areas increase unnecessary computational burden. Furthermore, due to information gaps at slice boundaries, local detection processes are independent of each other; for example, openings and defects in the ground area may be incorrectly associated with the ceiling area, forming false detection boxes with contradictory location logic, affecting the reliability of the detection results.

[0057] If the above issues are not addressed, the overall performance of the target detection system will be continuously affected. Decreased detection efficiency leads to delayed response times for inspection tasks, failing to meet real-time requirements; increased false positive rates reduce the reliability of detection results, potentially leading to erroneous maintenance decisions. In critical infrastructure inspections, false positives due to logically contradictory locations may result in undetected security risks, threatening the safe operation of facilities. Furthermore, inefficient allocation of computing resources limits the system's scalability in complex scenarios, making it difficult to adapt to diverse inspection needs.

[0058] Based on this, embodiments of this application provide a target detection method, apparatus, device, and medium. Through a mechanism of dynamic adaptive slicing and associated slice information, it can adapt to various panoramic image slicing scenarios and improve the quality of panoramic image slicing.

[0059] Figure 1 This diagram illustrates the application environment of the target detection method provided in the embodiments of this application. (See attached diagram.) Figure 1 This method is applied to a target detection system. The system includes a terminal 110 and a server 120. The terminal 110 and server 120 are connected via a network. The terminal 110 can be a desktop terminal or a mobile terminal; the mobile terminal can be at least one of a mobile phone, tablet, or laptop. The server 120 can be a standalone server or a server cluster consisting of several servers. The terminal 110 sends the panoramic image to be detected to the server 120. The server 120 acquires the panoramic image to be detected, performs adaptive slicing based on key image features in the panoramic image, obtains several original slice images, supplements the original slice images with neighborhood information, obtains several supplementary slice images, performs target detection on the supplementary slice images to generate corresponding target detection boxes, performs global consistency verification on the target detection boxes to remove target detection boxes that do not meet preset consistency conditions, and obtains the target detection result corresponding to the panoramic image to be detected.

[0060] It should be understood that Figure 1 The application scenarios shown are merely examples. In practical applications, the target detection method provided in this application embodiment can also be applied to other scenarios. For example, the above-described target detection method can be directly applied to terminal 110. Terminal 110 is used to acquire a panoramic image to be detected, and based on the key image features in the panoramic image to be detected, adaptively slices the panoramic image to be detected to obtain several original slice images. Neighborhood information is supplemented to the original slice images to obtain several supplementary slice images. Target detection is performed on the supplementary slice images to generate corresponding target detection boxes. Global consistency verification is performed on the target detection boxes to remove target detection boxes that do not meet the preset consistency conditions, thereby obtaining the target detection result corresponding to the panoramic image to be detected.

[0061] See Figure 2 In one embodiment, a target detection method is provided. The execution subject of the method can be either a terminal or a server, including but not limited to steps S201 to S205.

[0062] Step S201: Obtain the panoramic image to be detected.

[0063] Step S202: Based on the key image features in the panoramic image to be detected, adaptive slicing is performed on the panoramic image to be detected to obtain several original slice images.

[0064] Step S203: Neighborhood information is supplemented to the original slice image to obtain several supplementary slice images.

[0065] Step S204: Perform object detection on the supplementary slice image to generate corresponding object detection boxes.

[0066] Step S205: Perform global consistency verification on the target detection boxes to remove target detection boxes that do not meet the preset consistency conditions, and obtain the target detection results corresponding to the panoramic image to be detected.

[0067] The core innovation of this application lies in adapting the image content complexity to an adaptive slicing strategy driven by key image features, combining neighborhood information to bridge the gaps in slice boundary information, and introducing a global consistency verification mechanism to ensure that the detection results conform to the overall logic of the panoramic image. This systematically solves the problems of low detection efficiency and false detection caused by high computational resource consumption, insufficient target feature extraction, and information fragmentation between slices in ultra-high resolution panoramic images in scenarios such as power inspection.

[0068] Acquiring a panoramic image to be detected refers to the process of extracting panoramic image data from an image source. This can be achieved by manually selecting an image file and reading its contents, or by directly acquiring image data through a scanner or other image acquisition devices, such as the original image file acquired by a panoramic camera in a power inspection scenario.

[0069] Adaptive slicing based on key image features in the panoramic image to be detected refers to the process of dynamically dividing image regions according to the inherent attributes of the image content. Key image features can be understood as quantitative indicators reflecting the complexity of image content, which can be achieved through image histogram distribution analysis, edge density calculation, or texture entropy evaluation. For example, the slice boundary can be determined by calculating the pixel gradient variance of local regions, or the slice size can be dynamically adjusted according to the brightness change rate of the image region, thereby avoiding target fragmentation or context loss problems caused by fixed-size slices. Therefore, neighborhood information supplementation of the original sliced ​​image refers to the operation of enhancing the missing context information at the boundaries of adjacent slices. This can be achieved by using pixel-level interpolation methods to perform information fusion in overlapping areas, or by mapping non-overlapping pixel regions of adjacent slices based on feature matching. For example, bilinear interpolation can be used to supplement transition pixels at slice edges, or optical flow methods can be used to estimate motion vectors between adjacent slices to supplement dynamic scene information.

[0070] Target detection in supplementary slice images can be achieved using traditional target detection algorithms such as HOG features combined with support vector machine models, or by applying region candidate networks for target localization, such as identifying the contour features of structural defects in industrial equipment inspection.

[0071] Global consistency verification of target detection boxes refers to the process of verifying the rationality of detection results based on the overall semantics of panoramic images. It can be performed by using prior knowledge of the spatial distribution of target categories for location verification, or by using statistical methods to check whether the size distribution of detection boxes meets the preset threshold. For example, for ground target categories, detection boxes located in the top area of ​​the image can be removed, or a reference range for target size can be established based on historical inspection data to filter out abnormal results.

[0072] In the object detection method, the panoramic image to be detected is first acquired and used as the basic data input for subsequent processing. Adaptive slicing is performed based on key image features in the panoramic image, allowing the slicing strategy to dynamically adjust according to image content features. This avoids large targets being fragmented or small targets failing to be detected due to missing context, while also reducing redundant slices in sparse target regions. Furthermore, neighborhood information supplementation is performed on the original sliced ​​image. By analyzing the feature correlations between adjacent slices and performing feature mapping, the slices complement each other's non-overlapping region information, effectively bridging information gaps at slice boundaries. Therefore, performing object detection on the supplemented sliced ​​image can more accurately capture target features and improve detection reliability. Finally, a global consistency check is performed on the generated object detection boxes. Abnormal detection results that do not meet preset conditions are eliminated by combining target category and location information, ensuring that the final output conforms to the overall semantic consistency of the panoramic image.

[0073] In power facility inspection scenarios, this method can be applied to panoramic images with a single pixel size of 10000×5000. Key image features include local texture features of insulator defects and global structural features of utility poles. Adaptive slicing dynamically determines the slice size based on the target size distribution. For example, smaller slice areas are configured for small defects with a diameter of 15-30 pixels to preserve the complete context, while larger slice areas are used for large structures with a size exceeding 100 pixels. During the neighborhood information supplementation process, the feature correlation between adjacent original slice images is analyzed. When the feature similarity meets preset conditions, non-overlapping region information is mutually supplemented through feature mapping. Global consistency verification eliminates false detection boxes detected in the ground area that should be in the air, based on target category information.

[0074] As a preferred implementation, the above process effectively solves the problems of high computational resource consumption, low detection efficiency, insufficient target feature extraction, target fragmentation, insufficient contextual information, information separation between slices, and false detection in ultra-high resolution panoramic image target detection by combining feature-driven adaptive processing, inter-slice information collaboration, and global logical verification. It avoids target compression and redundant slices caused by fixed-size slices, while ensuring the global logical consistency and reliability of the detection results.

[0075] In some of the embodiments described above in this application, adaptive slicing based on key image features was proposed to optimize the panoramic image slicing strategy. However, in its implementation, due to the lack of a systematic extraction and parameter mapping mechanism for key image features, the slicing parameters are not accurately determined, and it cannot dynamically adapt to differences in target size and distribution density. This easily leads to problems such as large targets being fragmented, small targets having insufficient contextual information, and feature overlap in dense target areas and redundant slices in sparse areas. In response, this application further proposes a method for adaptive slicing of the panoramic image to be detected.

[0076] See Figure 3 In one embodiment, the method for adaptive slicing of the panoramic image to be detected includes, but is not limited to, steps S301 to S303.

[0077] Step S301: Extract key features from the panoramic image to be detected to obtain the corresponding target of interest features, local region features, and global image features.

[0078] Step S302: Based on the mapping relationship between key image features and slice parameters, determine the adaptive slice parameters corresponding to the target of interest features, local region features, and global image features.

[0079] Step S303: Based on adaptive slicing parameters, generate several preliminary slice images.

[0080] Step S304: Remove duplicates and fine-tune the preliminary slice image to obtain the original slice image.

[0081] Key feature extraction refers to extracting multi-level feature information from the panoramic image to be detected. It can be achieved by using deep convolutional neural networks or traditional edge detection algorithms. The purpose is to comprehensively obtain the size, position and contextual features of the target, so as to provide multi-dimensional data support for the subsequent determination of slicing parameters.

[0082] The mapping relationship between key image features and slice parameters can be understood as a dynamic transformation mechanism between key image features and slice parameters. It can be implemented using a predefined set of rules or a lightweight regression model. The purpose is to adjust the slice size and position in real time according to feature changes to ensure that the slicing strategy matches the image content features.

[0083] Generating preliminary slice images refers to segmenting panoramic images based on adaptive slicing parameters. This can be achieved using image cropping or affine transformation techniques, with the aim of creating initial slices that cover the target area and reducing the inclusion of invalid regions.

[0084] Deduplication and fine-tuning refer to the optimization processing of the initial sliced ​​image. This can be achieved by using a slice merging algorithm based on target semantics or a dynamic boundary adjustment strategy. The purpose is to eliminate redundant slices and ensure the integrity of the target, avoiding target fragmentation caused by slice boundary cutting.

[0085] The solution in this application extracts multi-scale features from panoramic images through key feature extraction, and then uses mapping relationships to convert the features into adaptive tiling parameters. These parameters guide the generation of preliminary tiled images. Based on this, deduplication and fine-tuning steps further optimize the tiling quality, ensuring that the tiles closely surround the target area and avoid fragmentation. The entire process forms a closed-loop adaptive mechanism, enabling the tiling strategy to dynamically respond to changes in image content, effectively solving the adaptation problem caused by fixed sizes.

[0086] As a preferred embodiment, the key feature extraction of the panoramic image to be detected is specifically implemented as follows: A lightweight object detection model (YOLO-Nano) is used to quickly scan the panoramic image to be detected, and outputs the coarse position and size information of the target to be detected. The semantic segmentation model MobileNet-UNet is used to divide the panoramic image to be detected into regions, distinguishing the target region, background region, and transition region. A global feature encoder (such as the global pooling layer of ResNet) is used to extract key features, and obtain the corresponding target of interest features, local region features, and global image features.

[0087] As a preferred embodiment, the adaptive slicing parameters corresponding to the target of interest features, local region features, and global image features are determined based on the mapping relationship between key image features and slicing parameters. The specific implementation is as follows: A regression model is used to predict the adaptive slicing parameters. The input to the regression model is the stitched target of interest features (such as target center coordinates, width and height), local region features (region complexity score), and global image features. A fully connected network (FCN) is used to learn the mapping relationship between key image features and slicing parameters through multi-layer nonlinear transformations, predicting the coordinate information of each preliminary slice image, thereby obtaining the corresponding adaptive slicing parameters. The training data uses manually labeled slice samples and their corresponding content features, and the loss function uses IoU loss and smoothed L1 loss.

[0088] As a preferred embodiment, the solution of this application is implemented as follows: In a power line inspection scenario, a ResNet-50 network is used to extract key features from the panoramic image to be inspected. The features of the target of interest focus on transformers and insulators, local region features capture local textures, and global image features describe the overall layout. The mapping relationship between key image features and slicing parameters is established using a random forest model to map features to slicing parameters. After generating preliminary slice images, an appropriate overlap rate is configured by analyzing the target size, and the slice positions are fine-tuned to ensure the transformer is fully contained within the slice. Invalid preliminary slice images containing only the sky area are discarded.

[0089] Through the above-described scheme, the embodiments of this application achieve accurate adaptive tiling of the panoramic image to be detected, effectively avoiding the problems of large targets being cut into fragments and small targets having insufficient contextual information. At the same time, it reduces feature overlap in dense target areas and redundant tiling in sparse areas, thereby improving the overall efficiency and accuracy of target detection.

[0090] In some embodiments, deduplication and fine-tuning of the initial slice image includes:

[0091] Based on the size characteristics of the target to be detected, the overlap rate between the initial slice images is configured;

[0092] When the edge of the preliminary slice image passes through the target to be detected, the position of the preliminary slice image is finely adjusted so that the preliminary slice image contains the complete target to be detected;

[0093] Remove the initial slice image containing only invalid regions.

[0094] Configuring overlap rate refers to dynamically adjusting the overlap ratio of adjacent slice regions according to the actual size of the target to be detected. It can be achieved by using a classification strategy based on size threshold or a scaling function. The purpose is to adapt to the slicing requirements of targets of different scales and avoid fragmentation of large targets due to insufficient overlap or dilution of context information due to excessive overlap of small targets.

[0095] Fine-tuning the position of the initial slice image can be understood as adjusting the coordinate range of the slice region by using a boundary offset algorithm when the slice boundary intersects with the target contour. This can be achieved based on edge detection results combined with target contour analysis. The purpose is to ensure that the target is completely contained within a single slice and to prevent feature truncation caused by fixed slice boundaries.

[0096] The solution in this application first dynamically configures the overlap rate between preliminary slice images based on the size characteristics of the target to be detected, ensuring that the slicing parameters match the target scale. This guarantees sufficient overlap for large targets to maintain structural continuity, while avoiding excessive overlap for small targets to preserve effective context. Subsequently, when the edge of the preliminary slice image passes through the target to be detected, the slicing region is fine-tuned by analyzing the target boundary position in real time, fully incorporating the target into the slicing range and resolving the cutting problem caused by fixed slicing strategies. Finally, invalid slices are filtered and removed based on the validity of the region content, reducing the amount of redundant data in subsequent processing. These steps are executed sequentially, forming a closed-loop optimization process that ensures the slicing process closely adheres to the image content features, effectively bridging the technical gap between local slicing and global target integrity.

[0097] As a preferred embodiment, the overlap rate between preliminary slice images is configured based on the size characteristics of the target to be detected as follows: Based on the size characteristics of the target to be detected, the target to be detected is clustered into several target clusters by k-means clustering, and then a corresponding overlap rate is configured for each target cluster, thereby configuring the overlap rate between preliminary slice images.

[0098] As a preferred embodiment, the solution of this application is implemented as follows: When processing a panoramic image containing transformers and insulators in a power inspection scenario, the execution entity configures a high overlap rate based on the larger size characteristics of the transformer to ensure that its structure is continuously presented in adjacent preliminary slice images. When the boundary of the preliminary slice image passes through the edge of the insulator, the slice area is automatically expanded to completely cover the insulator outline. At the same time, blank slice areas containing only the sky background are removed, so that inspection resources are concentrated in areas with dense equipment.

[0099] Through the above technical solutions, the embodiments of this application effectively solve the technical problems of targets being cut into fragments or lacking contextual information, ensure the feature integrity of targets of different scales in the slicing process, and significantly reduce redundant computation by removing invalid regions, thereby improving the accuracy and efficiency of target detection in panoramic images.

[0100] In some embodiments of this application, neighborhood information supplementation of the original slice image is proposed to enhance the information correlation between slices. However, in its implementation, due to the lack of a precise matching mechanism based on target contour features, the target contours of adjacent slices cannot be effectively correlated, resulting in the loss of target information in non-overlapping areas, which in turn leads to false detection and missed detection problems during target detection. In response, this application further proposes a method for supplementing neighborhood information of a pair of original slice images.

[0101] See Figure 4In one embodiment, the method for supplementing neighborhood information of a pair of original slice images includes, but is not limited to, steps S401 to S403.

[0102] Step S401: Extract the target contour features of the target to be detected from the original slice image.

[0103] Step S402: Calculate the contour similarity between the target contour features of adjacent original slice images.

[0104] Step S403: When the contour similarity meets the preset contour similarity conditions, feature mapping is performed on the target contour features so that the non-overlapping target contour features of the target to be detected are supplemented between adjacent original slice images to obtain supplementary slice images.

[0105] Target contour features refer to the boundary shape information of the target to be detected. They can be implemented using edge detection algorithms (such as the Canny operator) or deep learning-based contour prediction models. The purpose is to provide an accurate data foundation for contour similarity calculation and ensure the integrity of target boundary information.

[0106] Contour similarity can be understood as an indicator that quantifies the degree of matching of target contours in adjacent slices. It can be implemented using shape similarity measurement methods such as Hausdorff distance and Frechet distance. The purpose is to determine whether targets in adjacent slices belong to the same entity by thresholding, thus avoiding the erroneous association of unrelated contours.

[0107] Feature mapping specifically refers to the process of dynamically integrating the contour information of non-overlapping regions into adjacent slices. It can be achieved by feature point matching and affine transformation in image stitching technology. The purpose is to ensure that each slice contains a continuous and complete target contour, providing sufficient contextual information for subsequent detection.

[0108] The solution in this application first extracts the target contour features of the target to be detected from the original slice image to obtain the boundary shape information of the target. Then, it calculates the contour similarity between the target contour features of adjacent original slice images to quantify the degree of contour matching. When the contour similarity meets a preset condition, it triggers feature mapping of the target contour features, enabling adjacent original slice images to supplement each other with non-overlapping target contour features, thereby obtaining supplementary slice images. This process, based on the continuity judgment of contour features, dynamically integrates non-overlapping information at the slice boundary, effectively eliminating detection errors caused by local information fragmentation, while avoiding interference from irrelevant contours, ensuring the targetedness and accuracy of neighborhood information supplementation.

[0109] As a preferred embodiment, the solution of this application is implemented as follows: In a power line inspection scenario, for insulator detection, the execution entity extracts the contour features of the insulator from the original slice image. These contour features are used to obtain the closed boundary of the target through an edge detection algorithm. Subsequently, the similarity of the insulator contours in adjacent original slice images is calculated. If the similarity meets a preset condition, the contour information of non-overlapping areas is mapped to adjacent slices through feature point matching, ensuring that each slice contains a complete insulator contour structure. The resulting supplementary slice image can accurately identify insulators at the boundary during the target detection stage, avoiding misjudgments caused by contour fragmentation.

[0110] Through the above solution, the embodiments of this application effectively solve the problem of false detection and missed detection caused by the fragmentation of the target contour at the slice boundary, significantly improve the accuracy of target detection in panoramic images, and are especially suitable for the accurate identification of equipment faults in power inspection scenarios, ensuring the reliability and integrity of the detection results.

[0111] In some embodiments described above in this application, neighborhood information supplementation of the original slice image is proposed to enhance the information coherence between slices. However, in its implementation, the supplementation mechanism does not consider the similarity of texture features of adjacent slices, resulting in inaccurate supplemented information and causing false detections or missed detections during target detection. Therefore, this application further proposes another method for supplementing neighborhood information of the original slice image.

[0112] See Figure 5 In one embodiment, another method for supplementing neighborhood information of the original slice images includes, but is not limited to, steps S501 to S503.

[0113] Step S501: Extract overlapping and non-overlapping texture features between adjacent original slice images.

[0114] Step S502: Calculate the overlapping texture similarity between overlapping texture features of adjacent original slice images.

[0115] Step S503: When the overlapping texture similarity meets the preset texture similarity condition, feature mapping is performed on the non-overlapping texture features so that the non-overlapping texture features of adjacent original slice images can be supplemented to obtain supplementary slice images.

[0116] Overlapping texture features refer to the texture information presented in the overlapping area of ​​adjacent original slice images. They can be extracted using local binary mode or gray-level co-occurrence matrix, with the aim of identifying the texture consistency of adjacent slices in the shared area.

[0117] Overlapping texture similarity is an index that quantifies the degree of texture matching in overlapping regions. It can use cosine similarity or Euclidean distance to calculate the similarity between feature vectors, with the aim of dynamically determining whether adjacent slices belong to the same continuous scene.

[0118] The solution in this application extracts overlapping and non-overlapping texture features between adjacent original slice images, providing basic data for similarity calculation to distinguish shared and exclusive regions. Based on this, the overlapping texture similarity between overlapping texture features is calculated to dynamically assess whether adjacent slices belong to the same scene. When the similarity meets a preset condition, feature mapping of non-overlapping texture features is triggered, allowing adjacent slices to complement each other's non-overlapping features. This ensures that information supplementation only occurs when textures are highly similar, effectively preventing erroneous associations in areas of abrupt scene changes and avoiding false detections caused by blind supplementation.

[0119] As a specific implementation method, the solution of this application embodiment is implemented in the power inspection scenario as follows: When processing panoramic image slices, the system extracts the texture features of the overlapping areas of adjacent original slice images and calculates the similarity; if the similarity is higher than the threshold, the equipment contour features of the non-overlapping areas are mapped to the adjacent slices, so that each slice contains more complete power equipment context information, such as ensuring the continuity of equipment edges between slices when detecting transformers.

[0120] Through the above technical solutions, the embodiments of this application realize a dynamic evaluation mechanism based on texture feature similarity, which ensures the accuracy of neighborhood information supplementation, effectively reduces the risk of false detection of targets due to local information fragmentation, and improves the reliability of target detection in panoramic images.

[0121] In one embodiment, a method for supplementing neighborhood information of an original slice image includes:

[0122] Extract the target contour features of the target to be detected from the original slice image;

[0123] Calculate the contour similarity between target contour features in adjacent original slice images;

[0124] Extract overlapping and non-overlapping texture features between adjacent original slice images;

[0125] Calculate the overlapping texture similarity between overlapping texture features of adjacent original slice images;

[0126] When the contour similarity meets the preset contour similarity conditions, feature mapping is performed on the target contour features so that the non-overlapping target contour features of the target to be detected are mutually supplemented between adjacent original slice images. When the overlapping texture similarity meets the preset texture similarity conditions, feature mapping is performed on the non-overlapping texture features so that the non-overlapping texture features of the target to be detected are mutually supplemented between adjacent original slice images, resulting in supplementary slice images.

[0127] In some embodiments of this application, a global consistency check is proposed to eliminate target detection boxes that do not meet preset consistency conditions. However, in its implementation, the check process lacks targeted application of target category information, resulting in the inability to perform refined verification based on the actual category characteristics of the target (such as the small circular features of opening targets and the large elongated features of edge targets). This makes the size and shape check conditions too general, failing to effectively identify and eliminate false detection boxes caused by mismatched category features, such as misdetecting a ground opening as a non-target object in the ceiling area. In response, this application further proposes a method for performing global consistency check on a pair of target detection boxes.

[0128] See Figure 6 In one embodiment, the method for performing global consistency verification on a pair of target detection boxes includes, but is not limited to, steps S601 to S602.

[0129] Step S601: Obtain the target category information of the target to be detected.

[0130] Step S602: Based on the target category information, perform size verification and shape verification on the target detection boxes, and remove target detection boxes whose size verification results do not meet the preset size verification conditions and whose shape verification results do not meet the preset shape verification conditions.

[0131] Obtaining the target category information of the target to be detected means identifying the category type to which the target belongs. This can be achieved by using the category label output by the target detection model or a category recognition algorithm based on feature matching. The purpose is to provide category-specific prior knowledge for the verification process, ensuring that subsequent verification can dynamically adjust the verification criteria based on the inherent attributes of the target.

[0132] Based on target category information, size and shape verification of target detection boxes refers to dynamically setting verification parameters according to the target category. This can be achieved by matching and verifying category-related size threshold ranges and shape feature templates. The purpose is to ensure that the verification process closely matches the actual physical characteristics of the target, avoiding insufficient adaptation of general verification conditions to diverse targets. Removing target detection boxes whose size verification results do not meet preset size verification conditions and whose shape verification results do not meet preset shape verification conditions refers to removing invalid detection results that contradict category features. This can be achieved automatically through setting verification rules, aiming to directly eliminate false detections caused by category feature mismatch and improve the reliability of overall detection results.

[0133] The solution in this embodiment first obtains the target category information of the target to be detected, which serves as the benchmark for verification. Then, based on the target category information, size and shape verification conditions that match the category characteristics are dynamically generated. For example, a narrow size threshold is set for small-sized opening targets, focusing on circular contour verification; a wide size threshold is set for large-sized adjacent targets, emphasizing long strip feature matching. Next, size and shape verification are performed sequentially for each target detection box to ensure that the verification criteria strictly correspond to the target category. Finally, target detection boxes that fail both verifications are automatically removed. This process, through the adaptive adjustment of verification conditions driven by category information, achieves a more precise verification mechanism and effectively solves the problem of false detections caused by vague verification conditions.

[0134] As a specific implementation method, the solution of this application is implemented as follows: In a power inspection scenario, when the target category information is identified as "opening," the size verification uses a narrow threshold range to exclude excessively large detection frames, and the shape verification verifies the closure of the target contour through circular contour feature matching. When the target category information is identified as "near edge," the size verification allows a larger size range, and the shape verification verifies the linear extension characteristics of the target contour through a long strip feature template. This implementation method ensures that the verification process closely matches the actual physical properties of the target. For example, when inspecting the ceiling area, if the size of the target detection frame exceeds the typical range of the opening category or the shape deviates from the circular feature, it is automatically rejected, thereby avoiding misjudging non-target objects as openings.

[0135] Through the above technical solutions, the embodiments of this application realize refined global consistency verification for different types of targets, effectively eliminate false detection boxes caused by mismatch of category features, and significantly improve the accuracy and reliability of target detection results. It is especially suitable for accurate identification of diverse targets such as openings and edges in scenarios such as power inspection.

[0136] In some embodiments described above, a global consistency check is proposed to eliminate target detection boxes that do not meet the consistency conditions. However, in its implementation, the lack of a specific verification mechanism for the target position and spatial information leads to the inability to effectively identify and eliminate target detection boxes with positional deviations. For example, a hole that should be located on the ground is detected in the ceiling area, resulting in false detections. To address this, embodiments of this application further propose another method for global consistency check of target detection boxes.

[0137] See Figure 7 In one embodiment, another method for performing global consistency verification on a pair of target detection boxes includes, but is not limited to, steps S701 to S702.

[0138] Step S701: Obtain the target location information and target spatial information of the target to be detected.

[0139] Step S702: Based on the target location information and target spatial information, perform position verification on the target detection boxes and remove target detection boxes that deviate from the position represented by the target location information and the region represented by the target spatial information.

[0140] Target location information refers to the physical location description of the target to be detected in the scene. It can be implemented using a coordinate system or relative positional relationship, with the purpose of providing a benchmark reference for location verification.

[0141] Target spatial information refers to the area in space where the target to be detected should exist. It can be achieved using predefined regions or spatial constraint models, with the aim of defining the reasonable distribution area of ​​the target.

[0142] Position verification refers to the process of comparing the position of the target detection box with the preset position information. It can be achieved through coordinate matching or region inclusion checks, with the aim of identifying and eliminating detection results with abnormal position logic.

[0143] The solution in this application first obtains the target location information and target spatial information of the target to be detected as a verification benchmark. This information is established based on the physical distribution pattern of the target in the scene. Subsequently, the target detection box is positionally verified based on the target location information and target spatial information. By dynamically comparing the actual coordinates of the detection box with the preset position and area range, detection results with abnormal position logic are accurately identified. Finally, target detection boxes that deviate from the position represented by the target location information and the area represented by the target spatial information are removed, thereby ensuring that the detection results conform to the actual spatial constraints and forming a complete verification closed loop.

[0144] As a specific implementation method, the solution of this application embodiment is implemented as follows: In the scenario of building facility inspection, for the detection of opening targets, the system pre-configures the target location information and target spatial information that the opening should be located in the ground area; when the panoramic image is subjected to target detection, if a target detection box is generated in the ceiling area, the position verification module will detect that the actual position of the detection box deviates from the position represented by the ground position information and the area represented by the ground area information, thereby eliminating it and avoiding misdetecting the ceiling area as an opening.

[0145] Through the above solution, the embodiments of this application effectively eliminate target detection boxes with abnormal location logic, such as excluding detection results that appear in non-target areas, and significantly improve the global consistency and reliability of target detection.

[0146] In one embodiment, a global consistency check is performed on the target detection bounding box, including:

[0147] Obtain target category information, target location information, and target spatial information of the target to be detected;

[0148] Based on target category information, the target detection boxes are subjected to size and shape verification. Target detection boxes whose size verification results do not meet the preset size verification conditions and whose shape verification results do not meet the preset shape verification conditions are removed. Based on target location information and target spatial information, the target detection boxes are subjected to position verification. Target detection boxes that deviate from the position represented by the target location information and the region represented by the target spatial information are removed.

[0149] See Figure 8 This application also provides a target detection device that can implement the above-described target detection method. The device includes:

[0150] The first module 801 is used to acquire the panoramic image to be detected;

[0151] The second module 802 is used to adaptively slice the panoramic image to be detected based on key image features in the panoramic image to be detected, and obtain several original slice images.

[0152] The third module 803 is used to supplement the original slice image with neighborhood information to obtain several supplementary slice images;

[0153] The fourth module 804 is used to perform object detection on the supplementary slice image to generate corresponding object detection boxes;

[0154] The fifth module 805 is used to perform global consistency verification on the target detection boxes to remove target detection boxes that do not meet the preset consistency conditions, and obtain the target detection results corresponding to the panoramic image to be detected.

[0155] The specific implementation of this target detection device is basically the same as the specific implementation of the target detection method described above, and will not be repeated here.

[0156] Figure 9 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application.

[0157] The following reference Figure 9 To describe an electronic device 900 according to such an embodiment of the present disclosure. Figure 9 The electronic device 900 shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments disclosed herein.

[0158] like Figure 9 As shown, the electronic device 900 is presented in the form of a general-purpose computing device. The components of the electronic device 900 may include, but are not limited to: at least one processing unit 910, at least one storage unit 920, a bus 930 connecting different system components (including storage unit 920 and processing unit 910), a display unit 940, etc.

[0159] The storage unit stores program code, which can be executed by the processing unit 910, causing the processing unit 910 to perform the steps described in the above-described target detection method section of this specification according to various exemplary embodiments of this disclosure.

[0160] Storage unit 920 may include readable media in the form of volatile storage units, such as random access memory (RAM) 9201 and / or cache memory 9202, and may further include read-only memory (ROM) 9203.

[0161] Storage unit 920 may also include a program / utility 9204 having a set (at least one) program module 9205, such program module 9205 including but not limited to: operating system, one or more application programs, other program modules and program data, each or some combination of these examples may include an implementation of a network environment.

[0162] Bus 930 can represent one or more of several types of bus structures, including a memory cell bus or memory cell controller, a peripheral bus, a graphics acceleration port, a processing unit, or a local bus using any of the various bus structures.

[0163] Electronic device 900 can also communicate with one or more external devices 900' (e.g., keyboard, pointing device, Bluetooth device, etc.), and with one or more devices that enable a user to interact with electronic device 900, and / or with any device that enables electronic device 900 to communicate with one or more other computing devices (e.g., router, modem, etc.). This communication can be performed via input / output (I / O) interface 950. Furthermore, electronic device 900 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 960. Network adapter 960 can communicate with other modules of electronic device 900 via bus 930. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with electronic device 900, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0164] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0165] The target detection method, apparatus, device, and medium provided in this application avoid target fragmentation and context loss caused by fixed slices by adaptive slicing based on key image features and adapting content features. By supplementing with neighborhood information to establish an inter-slice association mechanism, it effectively alleviates the false detection problem caused by the fragmentation of boundary information. Combined with global consistency verification to verify the rationality of the detection results, it significantly improves the reliability of the detection results. It has the advantages of improving the accuracy and efficiency of target detection, avoiding target fragmentation and context loss, and reducing false detection and invalid slices. It can adapt to various panoramic image slicing scenarios and improve the quality of panoramic image slicing.

[0166] From the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, or network device, etc.) to execute the methods described above according to the embodiments of this disclosure.

[0167] The program product may employ any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of readable storage media include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0168] Computer-readable storage media may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable storage medium may also be any readable medium other than a readable storage medium that can transmit, propagate, or transfer a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.

[0169] Those skilled in the art will understand that the above modules can be distributed in the device as described in the embodiments, or they can be modified accordingly and placed in one or more devices that are unique to this embodiment. The modules in the above embodiments can be combined into one module, or they can be further divided into multiple sub-modules.

[0170] Exemplary embodiments of this disclosure have been specifically shown and described above. It should be understood that this disclosure is not limited to the detailed structures, arrangements, or implementations described herein; rather, this disclosure is intended to cover various modifications and equivalent arrangements contained within the spirit and scope of the appended claims.

Claims

1. A target detection method, characterized in that, include: Acquire the panoramic image to be detected; Based on the key image features in the panoramic image to be detected, adaptive slicing is performed on the panoramic image to be detected to obtain several original slice images. The original slice image is supplemented with neighborhood information to obtain several supplementary slice images; Target detection is performed on the supplementary slice image to generate corresponding target detection boxes; A global consistency check is performed on the target detection boxes to remove target detection boxes that do not meet the preset consistency conditions, thereby obtaining the target detection results corresponding to the panoramic image to be detected. The process of supplementing the original slice image with neighborhood information includes: Extract the target contour features of the target to be detected from the original slice image; Calculate the contour similarity between the target contour features of adjacent original slice images; When the contour similarity meets the preset contour similarity condition, feature mapping is performed on the target contour features so that the non-overlapping target contour features of the target to be detected are mutually supplemented between adjacent original slice images, thereby obtaining the supplementary slice image; The global consistency check of the target detection box includes: Acquire the target location information and target spatial information of the target to be detected; Based on the target location information and the target spatial information, the target detection box is positionally verified, and target detection boxes that deviate from the position represented by the target location information and the region represented by the target spatial information are removed.

2. The target detection method according to claim 1, characterized in that, The adaptive slicing of the panoramic image to be detected based on key image features includes: Key features are extracted from the panoramic image to be detected to obtain the corresponding target of interest features, local region features, and global image features; Based on the mapping relationship between the key image features and the slicing parameters, adaptive slicing parameters corresponding to the target of interest features, the local region features, and the global image features are determined; Based on the adaptive slicing parameters, several preliminary slice images are generated; The initial slice image is deduplicated and fine-tuned to obtain the original slice image.

3. The target detection method according to claim 2, characterized in that, The process of deduplication and fine-tuning the preliminary slice image includes: Based on the size characteristics of the target to be detected, the overlap rate between the initial slice images is configured; When the edge of the preliminary slice image passes through the target to be detected, the position of the preliminary slice image is finely adjusted so that the preliminary slice image contains the complete target to be detected; Remove the initial slice image containing only invalid regions.

4. The target detection method according to claim 1, characterized in that, The process of supplementing the original slice image with neighborhood information includes: Extract overlapping and non-overlapping texture features between adjacent original slice images; Calculate the overlapping texture similarity between overlapping texture features of adjacent original slice images; When the overlapping texture similarity meets the preset texture similarity condition, feature mapping is performed on the non-overlapping texture features so that the non-overlapping texture features are mutually supplemented between adjacent original slice images to obtain the supplementary slice image.

5. The target detection method according to claim 1, characterized in that, The global consistency check of the target detection box includes: Obtain the target category information of the target to be detected; Based on the target category information, the target detection boxes are subjected to size and shape verification. Target detection boxes whose size verification results do not meet the preset size verification conditions and whose shape verification results do not meet the preset shape verification conditions are removed.

6. A target detection device, characterized in that, include: The first module is used to acquire the panoramic image to be detected; The second module is used to adaptively slice the panoramic image to be detected based on the key image features in the panoramic image to be detected, and obtain several original slice images. The third module is used to supplement the original slice image with neighborhood information to obtain several supplementary slice images; The fourth module is used to perform target detection on the supplementary slice image to generate corresponding target detection boxes; The fifth module is used to perform global consistency verification on the target detection boxes to remove target detection boxes that do not meet the preset consistency conditions, and obtain the target detection results corresponding to the panoramic image to be detected. The process of supplementing the original slice image with neighborhood information includes: Extract the target contour features of the target to be detected from the original slice image; Calculate the contour similarity between the target contour features of adjacent original slice images; When the contour similarity meets the preset contour similarity condition, feature mapping is performed on the target contour features so that the non-overlapping target contour features of the target to be detected are mutually supplemented between adjacent original slice images, thereby obtaining the supplementary slice image; The global consistency check of the target detection box includes: Acquire the target location information and target spatial information of the target to be detected; Based on the target location information and the target spatial information, the target detection box is positionally verified, and target detection boxes that deviate from the position represented by the target location information and the region represented by the target spatial information are removed.

7. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the target detection method according to any one of claims 1 to 5.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the target detection method according to any one of claims 1 to 5.

Images