Self-adaptive cross-scene target detection method

Abstract

The invention provides a self-adaptive cross-scene target detection method, which comprises five core technical links: firstly, constructing a general enhanced block data set, and fusing public and scene special data sets to form a three-stage training data set; performing staged transfer learning on the pre-training model based on the training data set, and sequentially optimizing the block adaptability, the general recognition capability and the scene special detection performance; dynamically loading optimal parameters through a scene classifier; calling target scale knowledge to adaptively adjust the sizes of the sub-blocks (reducing when the target density is high and increasing when the target density is low), and respectively adjusting the detected sub-blocks; and results are integrated by adopting an area inverse proportion weighted fusion algorithm, so that the weight of a small-size sub-block detection result is higher. The method breaks through the bottleneck of small target detection through dynamic partitioning, realizes cross-scene migration in combination with a scene adapter, and remarkably improves the positioning precision of the small target.

Description

Technical Field

[0001] This invention relates to the field of object detection, and more particularly to an adaptive cross-scene object detection method. Background Technology

[0002] Traditional object detection algorithms perform well in recognizing large objects, but their accuracy drops significantly when dealing with small targets such as traffic signs and individuals in crowds. This is mainly because small targets have a low pixel ratio, making feature extraction difficult, and high-resolution image processing is inefficient. Existing improvements only partially alleviate the problem through multi-scale detection and cannot systematically address the core bottleneck of insufficient cross-scene adaptability. Full-image inference ignores local details, while fixed-block strategies waste computational power when targets are sparse and are prone to missing key targets when they are dense. Traditional methods struggle to balance detection accuracy and resource efficiency. Summary of the Invention

[0003] This invention proposes an adaptive cross-scene target detection method, characterized by comprising: S1. The collected high-resolution images of multiple scenes are divided into standard-sized sub-blocks to form a general enhanced block dataset. The enhanced block dataset is then merged with the public dataset and the scene-specific dataset to form a three-level training set. S2. Based on the three-level training set, the pre-trained model is transferred in stages to optimize the block adaptability, general target recognition and scene-specific detection capabilities of the pre-trained model in turn to output a fine-tuned model. S3. Identify the input type of the image to be detected through a scene classifier, and dynamically load the optimal parameters of the corresponding scene in the fine-tuning model; S4. Based on the target scale distribution knowledge of the three-level training set, the area and overlap rate of the sub-blocks in the image to be detected are adaptively adjusted to generate variable-size sub-blocks, and the optimal parameters are used to perform parallel inference detection of the sub-blocks. S5. The weighted nonmaximum suppression algorithm is used to integrate the sub-block detection results and retain the high-confidence target location information.

[0004] In S2, after optimizing the scene-specific detection capability, the method further includes: dividing the fine-tuning model into multiple exclusive parameter subsets according to scene type and storing them in the cloud or local parameter library; in S3, dynamically loading the optimal parameters corresponding to the scene in the fine-tuning model specifically involves: selecting and loading the exclusive parameter subset matching the scene from the parameter library based on the output of the scene classifier.

[0005] S2 further includes: extracting the target scale distribution knowledge of the three-level training set and encoding it into metadata for embedding into the fine-tuning model; and calling the metadata when adaptively adjusting the sub-block area and overlap rate in S4.

[0006] The scene classifier in S3 is optimized using the three-level training set through joint training or transfer learning, and is synchronously iterated and updated during the phased transfer learning process in S2. The input data of the scene classifier includes scene labels from the general augmented block dataset and annotation information from the scene-specific dataset.

[0007] In S5, the weights of the weighted nonmaximum suppression algorithm are dynamically calculated based on the area of ​​the sub-blocks generated in S4. Smaller sub-blocks are given higher weights to improve integration accuracy. Meanwhile, the scene classifier in S3 is configured to detect unknown scenes and is backed up using general target recognition parameters.

[0008] Specifically, in S4, the adaptive adjustment of sub-block area and overlap rate involves generating variable-size sub-blocks based on the target density, where the sub-block area is reduced when the target density is high and increased when the target density is low; the generated target density data is used in the parallel inference detection of S4 to optimize the sub-block processing efficiency.

[0009] In S5, the weights of the weighted nonmaximum suppression algorithm are dynamically calculated based on the sub-block area. Specifically, the weights are inversely proportional to the sub-block area, and smaller sub-blocks are given higher weights. The generated weight rules are used in the retention of high-confidence target location information in S5.

[0010] In S2, each subset of dedicated parameters is stored independently and loaded in isolation through the scene adapter. The generated isolated parameters are used in the optimal parameter loading of S3.

[0011] In the phased transfer learning process, the scene classifier weights are updated synchronously and iteratively, and the resulting updated classifier is used for input type recognition in S3.

[0012] Specifically, in S4, calling the metadata involves: obtaining target scale distribution knowledge through the metadata query module, dynamically calculating the area and overlap rate of sub-blocks; the generated metadata query results are used in the generation of variable-size sub-blocks in S4 to achieve knowledge-driven adaptive adjustment.

[0013] This method overcomes the bottleneck of small target detection through a dynamic block-segmentation mechanism: it adaptively adjusts the size of sub-blocks based on target density, enabling high-density areas to focus on detail capture while avoiding resource redundancy in low-density areas; combined with a scene adapter for one-click parameter switching, it achieves seamless migration to scenarios such as traffic monitoring and industrial inspection; and through an area-inverse weighted fusion algorithm, it significantly improves the boundary localization accuracy of tiny targets. The overall solution optimizes high-resolution image processing efficiency, enhances the model's robustness across scenarios, and balances real-time performance with accuracy requirements. Attached Figure Description

[0014] Figure 1This is a flowchart illustrating an adaptive cross-scene target detection method proposed in this invention. Detailed Implementation

[0015] refer to Figure 1 This invention proposes an adaptive cross-scene target detection method, including: S1. The collected high-resolution images from multiple scenes are segmented into standard-sized sub-blocks to form a general enhanced block dataset. The enhanced block dataset is then merged with the public dataset and the scene-specific dataset to form a three-level training set.

[0016] In the specific implementation of S1, firstly, high-resolution raw images from various real-world application scenarios are acquired. These images cover visually diverse environments. The acquisition process ensures the diversity and representativeness of the image sources. Next, a standardized segmentation operation is performed on each high-resolution image. Segmentation uses a uniformly preset standard size specification (basic size 512×512 pixels, supporting dynamic expansion to the [256×256, 1024×1024] range). This operation divides the large-size original image into multiple regular sub-image blocks. All sub-blocks maintain the exact same pixel size. Necessary overlapping areas are allowed between adjacent sub-blocks during segmentation. This overlapping design is to avoid the abrupt cutting of key targets. After segmentation, a primary set of sub-blocks is formed. Then, data augmentation processing is performed on this set. Augmentation methods include, but are not limited to, rotation, flipping, brightness adjustment, and noise addition. Augmentation operations significantly improve the diversity and robustness of the sub-block samples. The augmented set of sub-blocks constitutes a general augmented block dataset. This dataset has universal characteristics across scenarios. Then, three levels of data sources are merged. The first level is the newly constructed general augmented block dataset. The second level introduces publicly available large-scale object detection datasets as public datasets (COCO / ImageNet, injecting general object features). The third level integrates specialized labeled datasets for specific application scenarios (such as traffic sign sets / pedestrian sets, catering to specific scenario requirements). These three datasets are merged into a unified training set in a structured manner. The training set consists of: Level 1 public dataset → Level 2 public + specialized datasets → Level 3 fused general block dataset (optimized for small object detection). This training set possesses three key characteristics: The general enhanced block dataset provides basic object recognition capabilities. The public dataset injects rich general object feature knowledge. The scene-specific dataset caters to the specific recognition needs of particular scenarios. Standard-sized sub-block segmentation overcomes the memory bottleneck of hardware processing high-resolution images. This segmentation allows images from different sources to achieve uniform processing specifications. The three-level fusion structure unexpectedly achieves a dynamic balance between generality and specialization, which traditional methods struggle to achieve. The general enhanced block dataset enables the model to learn cross-scene invariant features. The public dataset expands the breadth of target categories the model can recognize. The scene-specific dataset deepens the cognitive depth of specific scenarios. The synergistic effect of the three data levels produces a knowledge complementarity effect. This effect significantly improves the stability of subsequent cross-scene transfer. Preserving overlapping regions during segmentation avoids the loss of target information due to fragmentation. This design ensures the integrity of subsequent detection boundaries. Standardized sub-block sizes create the foundation for subsequent parallel processing. This processing architecture greatly improves overall efficiency. Data augmentation effectively simulates the complex variations of real-world scenarios. The augmented dataset significantly improves the model's robustness to interference. The three-level training set constitutes a hierarchical knowledge system. This system becomes the core support for subsequent adaptive detection.

[0017] S2. Utilizing a three-level training set, optimization is performed through joint training or transfer learning. During the phased transfer learning process on the pre-trained model, a scene classifier is iteratively updated. The input data for the scene classifier includes scene labels from a general augmented block dataset and annotation information from scene-specific datasets. Based on the three-level training set, the pre-trained model undergoes phased transfer learning. During this process, the scene classifier weights are iteratively updated, generating an updated classifier. The block adaptability of the pre-trained model is optimized; general object recognition is optimized; scene-specific detection capabilities are optimized, and a fine-tuned model is output. The fine-tuned model is divided into multiple scene-specific parameter subsets, which are then stored in the cloud or local parameter library. Each parameter subset is stored independently and can be loaded in isolation via a scene adapter, generating isolated parameters. The target scale distribution knowledge of the three-level training set is extracted and encoded as metadata, embedded into the fine-tuned model.

[0018] In the specific implementation of S2, a three-level training set is used to perform phased transfer learning on the pre-trained model (YOLOv7 / x). First, the block adaptability of the model is optimized through a general augmented block dataset to ensure that the model can efficiently process standard-sized sub-blocks and avoid the memory limitations caused by high-resolution images. Next, a public dataset is introduced to optimize the general object recognition capability, expand the range of objects the model can recognize, and enhance cross-scene generalization. Then, a scene-specific dataset is used to optimize the scene-specific detection capability, deepening the model's understanding of specific environments. In the phased transfer learning process, Level 1: Training on the public dataset → optimizing basic feature extraction; Level 2: Public + scene-specific datasets → enhancing scene generalization; Level 3: General block dataset → optimizing small object detection; the weights of the scene classifier are updated synchronously and iteratively. The input data combines the scene labels from the general augmented block dataset and the annotation information from the scene-specific dataset, enabling the classifier to capture scene features and generate an updated classifier. After outputting the fine-tuned model, it is divided into multiple dedicated parameter subsets according to scene type. Each dedicated parameter subset is independently stored in the cloud or local parameter library. One-click mode switching (transportation / security / infrastructure mode) is achieved through the scene adapter API. Isolated loading via the scene adapter ensures parameter independence. The output fine-tuned model must meet the real-time performance requirement of ≥30fps on edge devices. Simultaneously, the target scale distribution knowledge of the three-level training set is extracted and encoded as metadata embedded in the fine-tuned model to support subsequent adaptive adjustments. This phased transfer learning improves the overall adaptability and training efficiency of the model because synchronously updating the scene classifier reduces additional training overhead, thus achieving a dynamic balance between generality and specialization. Traditional methods often lead to performance fragmentation or resource waste due to independent optimization, but this method, through three-level data collaboration and parameter isolation, brings unexpected improvements in stability and robustness. Optimizing block adaptability accelerates image processing; enhancing general recognition reduces false detection rates; and optimizing specific detection improves accuracy for specific scenes. Dedicated parameter subset storage makes it lightweight and easy to deploy quickly. Metadata embedding lays the foundation for knowledge-driven adjustments. Overall, the S2 implementation shortens model fine-tuning time and enhances flexibility. Traditional methods struggle to accommodate the characteristics of different datasets, easily leading to overfitting or undergeneralization. However, this method overcomes information asymmetry through phased fusion and synchronous updates, achieving efficient knowledge transfer and reducing computational resource requirements. Parameter isolation mechanisms prevent interference between scenes, improving reliability. Scene classifier updates are performed synchronously with model optimization, reducing training iterations. Target-scale knowledge extraction enhances subsequent adaptive capabilities. Storing a dedicated subset of parameters supports parallel processing across multiple scenes. After fine-tuning the model output, parameter library management simplifies maintenance. This implementation improves cross-scene detection accuracy because the complementary effect of the three-level training sets strengthens the model's learning depth. Scene adapters ensure accurate parameter loading, avoiding incorrect matching. Metadata encoding endows the model with intrinsic tuning knowledge.Phased transfer learning balances the learning rate and prevents premature convergence. Updating the classifier enhances subsequent scene recognition capabilities. A dedicated subset of parameters is stored independently, supporting dynamic expansion. The overall process reduces deployment complexity; traditional optimization methods require multiple parameter tunings, but this method, through integrated optimization steps, unexpectedly achieves a one-stop solution, resulting in greater practicality and scalability.

[0019] S3. Identify the input type of the image to be detected by updating the classifier; based on the output of the scene classifier, select a subset of exclusive parameters matching the scene from the parameter library and load them to obtain isolated parameters so as to realize the dynamic loading of the optimal parameters of the corresponding scene in the fine-tuning model; wherein, the scene classifier is configured to detect unknown scenes and is backed up by general target recognition parameters.

[0020] In the specific implementation of S3, firstly, the image to be detected is input into the update classifier for scene type identification. The update classifier is obtained based on synchronous iterative training in the S2 stage, and its internal weights have integrated scene label features from the general augmented segment dataset and the labeled semantics from the scene-specific dataset. This classifier uses a multi-layer feature extraction network to parse deep information such as the global composition, illumination distribution, and texture patterns of the input image, generating a scene category probability vector. Next, the scene classifier determines the most likely scene type based on the probability vector and triggers a parameter library retrieval instruction. The parameter library stores multiple dedicated parameter subsets segmented in the S2 stage, each subset corresponding to optimized weights for a specific scene. The scene adapter accurately loads the dedicated parameter subset matching the scene based on the retrieval results, forming an isolated parameter module. The scene adapter loads parameter subsets based on the classification results and can also support real-time switching of preset scene modes (e.g., traffic mode → security mode). This module uses memory isolation technology to prevent parameter crosstalk, ensuring that the loaded weights only apply to the current scene inference. If the scene classifier detects unknown scene features, i.e., all scene probabilities are below the adaptive threshold, it automatically switches to the backup loading mode. The backup mode calls a pre-stored subset of general object recognition parameters, trained on a public dataset and possessing basic object recognition capabilities, or general object recognition parameters trained on a public dataset. During loading, the compatibility of parameters with the scene is verified in real time, and the validity of the parameters is confirmed through feature similarity calculation. After loading is complete, isolated parameters are immediately embedded into the fine-tuned model architecture, replacing the original full-scene parameters and achieving dynamic reconstruction of the model kernel. Through dynamic scene recognition and accurate parameter loading, the adaptability of cross-scene detection can be improved. Traditional methods suffer from feature perception bias due to the independent training of scene classifiers and detection models, often leading to incorrect parameter loading in complex equipment scenes and traffic monitoring scenes in industrial inspection. This method, however, synchronously updates the classifier through transfer learning, enabling scene feature understanding and detection capability optimization to co-evolve, thereby eliminating differences in feature expression. S3 reduces the risk of memory conflicts during multi-task parallel processing through the isolated storage design of the parameter library. The backup loading mechanism retains basic functions through general parameters. The isolated loading technology of the scene adapter improves security; when a scene parameter is contaminated, it will not affect other scene subsets. This design can effectively block model attacks targeting specific scenes. In cross-city traffic monitoring networks, this method enables seamless switching between highway and urban congestion scenarios. Traditional solutions require preloading all models, leading to memory overflows. By dynamically reconstructing the model kernel, this method allows edge devices to handle multiple scenario tasks with only a small parameter library, whereas traditional solutions require deploying multiple complete models. Ultimately, a closed-loop scenario adaptation mechanism is formed, which reduces the frequency of manual intervention through real-time interaction between the classifier and the parameter library.

[0021] S4. Based on the target scale distribution knowledge of the three-level training set, metadata is invoked, and the area and overlap rate of sub-blocks in the image to be detected are adaptively adjusted to generate variable-size sub-blocks. The generation of variable-size sub-blocks is based on the target density; when the target density is high, the sub-block area is reduced, and when the target density is low, the sub-block area is increased to optimize sub-block processing efficiency. Optimal parameters are used for parallel inference detection of sub-blocks, and the inference detection process can utilize target density data. Invoking the metadata specifically involves obtaining target scale distribution knowledge through the metadata query module and dynamically calculating the sub-block area and overlap rate. The generated metadata query results are used in the generation of variable-size sub-blocks to achieve knowledge-driven adaptive adjustment.

[0022] In the specific implementation of S4, an adaptive block segmentation mechanism is initiated by invoking the target scale distribution knowledge embedded in the fine-tuning model. First, the metadata query module parses the overall features of the image to be detected and extracts pre-stored target scale distribution knowledge, which originates from the statistical regularities of target sizes in different scenes within the three-level training set. Next, the sub-block area and overlap rate are dynamically calculated based on the target density of the current image. For example, target density ≥ 0.5 targets / pixel² (dense) → sub-block size 256×256; target density < 0.1 targets / pixel² (sparse) → sub-block size 1024×1024 (supplementary quantization rule). The optimal parameters are used for parallel inference of the sub-blocks via the GPU. When the target density is high, the sub-block area is automatically reduced and the overlap area is increased to ensure complete coverage of dense small targets; when the target density is low, the sub-block area is automatically increased and the overlap area is reduced to avoid resource redundancy. This dynamic adjustment directly generates variable-size sub-blocks that match the scene requirements. By responding to changes in target density in real time, the efficiency bottleneck caused by traditional fixed-block segmentation can be overcome. Traditional methods, unable to predict target distribution patterns, rely on a uniform block-splitting strategy. This results in numerous empty sub-blocks wasting computational power in sparse target scenarios, while excessively large blocks lead to missed detections of small targets in dense target scenarios. Our proposed method, however, utilizes knowledge-driven dynamic block-splitting to significantly improve processing speed in high-density scenarios while maintaining a high capture rate for small targets. In low-density scenarios, it reduces unnecessary computation and saves memory. After generating variable-size sub-blocks, pre-loaded isolation parameters are automatically invoked for parallel inference. The inference process simultaneously utilizes target density data to optimize detection focus, for example, enhancing the weight of the small target recognition layer in high-density regions. By fusing scale distribution knowledge, the block-splitting strategy becomes self-determined, eliminating the trial-and-error costs of manual parameter tuning. Traditional solutions require manual setting of block-splitting rules for different scenarios, while our method automatically adapts through an internal knowledge base, achieving optimal block-splitting schemes for both precision parts inspection in industrial settings and open farmland in agricultural monitoring. This adaptive mechanism also addresses the contradictory requirements of high-resolution image processing: large sub-blocks are beneficial for capturing global context but easily lose details, while small sub-blocks are excellent at target localization but ignore related information. By dynamically balancing the size of sub-blocks as needed, both detection accuracy and efficiency can be considered. Ultimately, the variable-size sub-blocks lay the foundation for subsequent weighted fusion, with their area parameters directly participating in the weight calculation for result integration, forming a closed-loop optimization chain.

[0023] S5. The sub-block detection results are integrated using a weighted nonmaximum suppression algorithm. The dynamic calculation rule of the weight of the weighted nonmaximum suppression algorithm is based on the sub-block area. The weight is inversely proportional to the sub-block area, and smaller sub-blocks are given higher weights. High-confidence target location information is retained. The dynamic calculation rule of the weight can be used in the process of retaining high-confidence target location information.

[0024] In the specific implementation of S5, firstly, variable-sized sub-blocks and their parallel inference results are received. Each sub-block contains target localization bounding boxes and confidence data. The core of the weighted nonmaximum suppression algorithm lies in the dynamic weight allocation mechanism, which uses the sub-block area as the calculation basis: smaller sub-blocks are assigned higher weights, and larger sub-blocks are assigned lower weights. The weight calculation module extracts the sub-block size parameters in real time and generates weight coefficients through an inverse proportional function. For example, in industrial inspection, small-sized sub-blocks generated due to the dense target area in the precision parts region automatically have higher weights than large-sized sub-blocks in the background region.

[0025] Next, the detection box fusion process begins: all overlapping detection boxes are scanned, and multiple detection boxes for the same target are sorted by weight. Smaller sub-block detection boxes with higher weights are retained first, while larger sub-block detection boxes with lower weights are suppressed. The weighting coefficients directly participate in the confidence weighting calculation, ensuring that the localization data of smaller sub-blocks dominates the final result. For example, in traffic monitoring, the tiny pedestrian target boxes at densely populated intersections, originating from high-weight sub-blocks, retain accurate boundaries after fusion.

[0026] S5 overcomes the accuracy limitations of traditional non-maximum suppression by using an inverse area weighting rule. Traditional methods treat all detection boxes equally; when large sub-blocks become blurred due to a wide field of view, their low-quality detection boxes interfere with the determination of high-precision small boxes. This method, however, prioritizes the detailed detection results of densely populated target areas through weight differentiation, significantly improving the aggregation accuracy of small targets. In agricultural pest and disease monitoring, traditional methods often result in insect egg locations being covered by background boxes because densely populated areas of insect eggs have the same weight as large areas of leaf background. S5, on the other hand, increases the weight of microscopic egg boxes in microscope images, ensuring that the fusion results do not lose key targets.

[0027] Meanwhile, the weighting mechanism and S4's variable-size block division form a closed-loop synergy. When S4 shrinks the sub-blocks according to the target density, S5 automatically assigns them high weights. This linkage ensures self-consistency in resource allocation: computational resources are tilted towards high-value areas. For example, in satellite image detection, urban building clusters have high weights due to their small blocks, while open suburban areas have low weights due to their large blocks. The final fusion result not only fully preserves the building outlines but also avoids interference from false alarm boxes in the suburbs.

[0028] Finally, only detection boxes with a weighted confidence score exceeding the adaptive threshold are retained, and the integrated target location and category are output. S5 uses dynamic weighting to suppress low-quality redundant boxes, reducing the false negative rate across scenes, and particularly improving the recall rate for densely packed small targets. Traditional methods in medical cell detection often obscure adjacent cells due to blurred boxes in large sub-blocks, while this method protects the fine detection of small sub-blocks with high weights, improving the success rate of separating and detecting adherent cells. This weighting design based on physical size unexpectedly solves the problem of unbalanced result reliability caused by variable-scale segmentation, resulting in a dual gain in accuracy and efficiency.

[0029] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An adaptive cross-scene target detection method, characterized in that, include: S1. The collected high-resolution images of multiple scenes are divided into standard-sized sub-blocks to form a general enhanced block dataset. The enhanced block dataset is then merged with the public dataset and the scene-specific dataset to form a three-level training dataset. S2. Based on the three-level training dataset, the pre-trained model is transferred in stages to optimize the block adaptability, general target recognition and scene-specific detection capabilities of the pre-trained model in turn to output a fine-tuned model. S3. Identify the input type of the image to be detected through a scene classifier, and dynamically load the optimal parameters of the corresponding scene in the fine-tuning model; S4. Based on the target scale distribution knowledge of the three-level training set, the area and overlap rate of the sub-blocks in the image to be detected are adaptively adjusted to generate variable-size sub-blocks, and the optimal parameters are used to perform parallel inference detection of the sub-blocks. S5. The weighted nonmaximum suppression algorithm is used to integrate the sub-block detection results and retain the high-confidence target location information.

2. The adaptive cross-scene target detection method as described in claim 1, characterized in that, In S2, after optimizing the scene-specific detection capability, the method further includes: dividing the fine-tuning model into multiple exclusive parameter subsets according to scene type and storing them in the cloud or local parameter library; in S3, dynamically loading the optimal parameters corresponding to the scene in the fine-tuning model specifically involves: selecting and loading the exclusive parameter subset matching the scene from the parameter library based on the output of the scene classifier.

3. The adaptive cross-scene target detection method as described in claim 2, characterized in that, S2 also includes: extracting the target scale distribution knowledge of the three-level training dataset and using it as a parameter for fine-tuning the model; the above parameters are called when adaptively adjusting the sub-block area and overlap rate in S4.

4. The adaptive cross-scene target detection method as described in claim 3, characterized in that, The scene classifier in S3 is optimized using the three-level training dataset through joint training or transfer learning, and is synchronously iterated and updated during the phased transfer learning process in S2. The input data for the scene classifier includes scene parameters from a general augmented block dataset and annotation information from a scene-specific dataset.

5. The adaptive cross-scene target detection method as described in claim 4, characterized in that, In S5, the weights of the weighted nonmaximum suppression algorithm are dynamically calculated based on the area of ​​the sub-blocks generated in S4. Smaller sub-blocks are given higher weights to improve integration accuracy. Meanwhile, the scene classifier in S3 is configured to detect unknown scenes and is backed up using general target recognition parameters.

6. The adaptive cross-scene target detection method as described in claim 5, characterized in that, In S4, the adaptive adjustment of sub-block area and overlap rate specifically involves generating variable-size sub-blocks based on the target density, wherein the sub-block area is reduced when the target density is high and increased when the target density is low; the generated target density data is used in the parallel inference detection of S4 to optimize the sub-block processing efficiency.

7. The adaptive cross-scene target detection method as described in claim 6, characterized in that, In S5, the weights of the weighted nonmaximum suppression algorithm are dynamically calculated based on the sub-block area. Specifically, the weights are inversely proportional to the sub-block area, and smaller sub-blocks are given higher weights. The generated weight rules are used in the retention of high-confidence target location information in S5.

8. The adaptive cross-scene target detection method as described in claim 2, characterized in that, In S2, each subset of dedicated parameters is stored independently and loaded in isolation through the scene adapter. The generated isolated parameters are used in the optimal parameter loading of S3.

9. The adaptive cross-scene target detection method as described in claim 4, characterized in that, During the phased transfer learning process, the scene classifier weights are updated synchronously and iteratively, and the resulting updated classifier is used for input type recognition in S3.

10. The adaptive cross-scene target detection method as described in claim 3, characterized in that, In S4, the metadata is invoked in the following ways: the target scale distribution knowledge is obtained through the metadata query module, and the sub-block area and overlap rate are dynamically calculated; the generated metadata query results are used in the variable-size sub-block generation in S4 to achieve knowledge-driven adaptive adjustment.

Images