Method for identifying target based on model complementation

Abstract

The invention relates to the technical field of reasoning models, and discloses a model complementation-based target recognition method, which comprises a picture preprocessing step, a target recognition model preliminary recognition step, a confidence coefficient filtering step, a large model scene complementation detection step and a multi-model result fusion step. According to the method, consistent image input is provided through standardization processing, feature extraction deviation caused by size and color differences is reduced, meanwhile, the target recognition model more systematically captures landslide features through grid division and feature extraction, and missed detection of features with tiny expressions is reduced; secondly, low-confidence misjudgment can be filtered through confidence threshold screening, and different feature requirements can be met; furthermore, secondary detection of the large model supplements the complex scene recognition capability, and missing detection and misjudgment of a single model under low confidence or rare geological scenes are reduced; in addition, the weight fusion mechanism integrates the advantages of the target recognition model and the large model, and the overall accuracy and robustness of landslide detection are improved.

Description

Technical Field

[0001] This application relates to the field of reasoning model technology, specifically a method for identifying targets based on model complementarity. Background Technology

[0002] In the field of modern computer vision, target recognition technology, as a core support for intelligent perception, has been widely applied in numerous scenarios such as security monitoring, autonomous driving, geological disaster monitoring, and industrial quality inspection. Among them, traditional target detection models, represented by YOLOv5, are widely used in scenarios with high real-time requirements, such as geological disaster monitoring (e.g., landslide and debris flow early warning), due to their high efficiency in real-time operation. The typical workflow of such models usually includes: performing basic preprocessing on geological images captured by UAVs or satellite remote sensing (e.g., uniformly scaling to a fixed size, normalizing pixel values ​​to the range of 0-1); decomposing the image into several sub-regions through a preset grid division strategy; extracting local visual features of each grid based on a convolutional neural network; determining whether there are signs of landslides within the grid, the type of landslide, and the bounding box; outputting the confidence level of the corresponding recognition result; and finally filtering valid results through a fixed threshold.

[0003] However, in geological morphology detection scenarios such as landslides, the limitations of existing technologies become prominent as terrain complexity and environmental interference (such as vegetation cover and weather effects) increase:

[0004] First, it has a weak ability to identify minute geological features and obscured scenes. Traditional models rely on local grid features for judgment. For minute landslide features that account for less than 1% of the image (such as surface cracks only a few centimeters wide in drone aerial images, or small staggered marks on mountain slopes), these features are often drowned out by background noise such as surface vegetation and rock shadows, resulting in a false negative rate of over 30%. Furthermore, when landslide signs are partially obscured (such as the rear edge of a landslide covered by dense shrubs, or the sliding surface covered by deposits), the model struggles to infer the complete landslide morphology from locally visible features, leading to a significant drop in accuracy.

[0005] Secondly, the geological category recognition has obvious limitations. The recognition ability of traditional models relies heavily on pre-trained datasets and can only accurately identify specific landslide types seen during training (such as common soil landslides and bedding landslides). They are almost unable to identify rare geological features that have not been trained (such as permafrost landslides in high-altitude areas, rock landslides caused by rock fracture zones, and volcanic debris landslides in volcanic activity areas), resulting in low recognition accuracy.

[0006] Third, traditional model preprocessing only reaches the standardization level and lacks adaptive optimization for complex geological scenarios: In heavy rain or fog, aerial images suffer from low contrast and blurred details due to rain and fog interference, and unenhanced features make it difficult to distinguish landslide traces (such as soil sliding zones) from the background mountains; in areas with dense small landslide points (such as secondary landslide groups after earthquakes), the lack of local feature magnification processing leads to the neglect of small landslides. At the same time, the fixed size and density of the mesh generation means that in mountainous scenarios with dense small landslide points (such as dozens of sporadic landslides triggered by a heavy rainfall), the fixed mesh cannot accurately capture the boundaries and features of each landslide; while in the vast transition zone between plains and mountains (where landslide signs are sparse), the redundant mesh results in a waste of computational resources.

[0007] Therefore, there is an urgent need for an optimization scheme that integrates the real-time performance of traditional models with the scene understanding capabilities of large models to improve image recognition accuracy and generalization ability. Summary of the Invention

[0008] The purpose of this application is to provide a method for identifying targets based on model complementarity, so as to solve the technical problems mentioned in the background art above.

[0009] To achieve the above objectives, this application discloses the following technical solution: a method for identifying targets based on model complementarity, the method comprising the following steps:

[0010] Image preprocessing steps: Standardize the input image, which includes adjusting the input image to a preset uniform size and normalizing the color values;

[0011] Preliminary identification steps of the target recognition model: The target recognition model divides the standardized image into grids and extracts features, identifies the existence, category, and location box of the target within each grid, and outputs the confidence level of the corresponding recognition result;

[0012] Confidence filtering step: Based on the set confidence threshold, retain the recognition results with a confidence level exceeding the threshold as valid preliminary results;

[0013] Large model scene complementarity detection step: When the valid preliminary result output by the confidence filtering step is empty or there are low-confidence candidate targets, the large model with the pre-configured geological scene knowledge base is called to perform secondary detection. The secondary detection includes the large model calling the corresponding multi-dimensional verification rules based on the geological scene type to jointly analyze the global features and local details of the image.

[0014] Multi-model result fusion step: Based on the weight allocation mechanism, the effective preliminary results are fused with the detection results of the large model, and a comprehensive recognition result is output; wherein, the weights are adjusted in real time according to the target category, scene complexity and historical model performance.

[0015] Preferably, the image preprocessing step further includes: performing adaptive enhancement operations to optimize image features in complex scenes while performing standardization processing;

[0016] The adaptive enhancement operation includes: adaptive denoising based on noise detection, and at least one of the following: contrast enhancement for low-light scenes, edge sharpening for complex backgrounds, and local feature magnification for small targets.

[0017] Preferably, in the confidence filtering step, the confidence threshold is adjusted according to the target category, including: using a threshold no lower than the high-priority target threshold parameter for preset high-priority targets, and using a threshold no lower than the regular target threshold parameter for regular targets, and the threshold can be iteratively updated through model training data and validation results.

[0018] Preferably, in the large model scene complementarity detection step, the large model includes a preset geological scene and its corresponding verification rule base; the multi-dimensional verification rules include:

[0019] Scene feature verification: Identify the distribution characteristics and proportion of large areas of exposed soil or gravel; detect whether the terrain has unnatural tilts, cracks, or stepped deformations;

[0020] Spatiotemporal comparison verification: Compare the current image with historical images to analyze changes in vegetation cover and soil displacement trajectory;

[0021] Semantic reasoning verification: Based on geological logic reasoning, inferring the morphological characteristics of potential landslide risk areas through regional geological data.

[0022] Preferably, the weight allocation mechanism includes weight allocation basis, weight adjustment rules, and weight optimization strategy;

[0023] The weight allocation is configured as follows: weights are determined based on target category, scene complexity, and model historical recognition accuracy; wherein, the target category includes known target categories that the target recognition model has trained on, and untrained target categories that the large model recognizes; the scene complexity includes whether the target is occluded, whether it is a small target, and the degree of background interference; the model historical recognition accuracy is the statistical value of the historical correct recognition rate of the target recognition model and the large model in the same type of target or the same scene;

[0024] The weight adjustment rules include: when the confidence level of the target recognition model for a known category of target is not lower than the high confidence parameter of the known category, and its historical recognition accuracy is not lower than the preset benchmark value, then the weight ratio of the effective preliminary results output by the target recognition model is not lower than the high weight parameter of the target recognition model; when the semantic reasoning confidence level of the large model for a new category of target or a target in a complex scene is not lower than the high confidence parameter of the new category, and its historical recognition accuracy is not lower than the preset benchmark value, then the weight ratio of the detection results output by the large model is not lower than the high weight parameter of the large model.

[0025] When there is a conflict between the output of the target recognition model and the larger model, the weights are tilted toward the model with higher historical recognition accuracy.

[0026] The weight optimization strategy includes: based on a reinforcement learning algorithm, using the error between the comprehensive recognition result and the manually labeled result as the reward signal, and periodically updating the parameters in the weight adjustment rule.

[0027] Preferably, in the preliminary recognition step of the target recognition model, the standardized image is divided into grids and its features are extracted, including the following steps:

[0028] The grid size and density are adjusted based on the target density and size distribution characteristics of the standardized image, including: dividing regions with small targets into high-density fine-grained grids, and dividing regions without targets or with large targets into low-density coarse-grained grids.

[0029] Based on the fusion of shallow low-dimensional edge features and deep high-dimensional semantic features, the feature weights of key regions of the target are enhanced through an attention mechanism to complete feature extraction.

[0030] Preferably, in the preliminary identification step of the target recognition model, when identifying the existence, category, and location box of the target in each grid, the target recognition model also performs correlation reasoning on the target features of adjacent grids through the constructed inter-grid feature correlation matrix.

[0031] Preferably, the correlation reasoning includes:

[0032] Determine the association type, which includes whether there is a spatial occlusion relationship between targets in adjacent grids, whether they belong to different parts of the same target, and whether they have a semantic logical relationship.

[0033] Based on the judgment result of the association type, the judgment of the target category and the coordinates of the location box are corrected. At the same time, the association confidence is calculated, and the confidence of the recognition result output by the target recognition model is corrected by the association confidence.

[0034] Preferably, the determination of the association confidence level includes the following steps:

[0035] Extract multi-dimensional features of targets in adjacent grids, including spatial location features, visual features, and semantic features;

[0036] Construct a feature association scoring matrix, assign feature weights to the extracted multi-dimensional features, and calculate the initial association score by weighted summation;

[0037] A correction factor is set based on the association type to calibrate the initial association score, resulting in the final association confidence score, where:

[0038] If targets within adjacent grids are determined to belong to different parts of the same target, the correction factor is dynamically adjusted based on the similarity of geological features: when the feature similarity is not less than 80%, the correction factor is not less than 0.7; when the feature similarity is between 30% and 80%, the expression for the correction factor is: correction factor = 0.3 + 0.8 × (feature similarity - 0.3); when the feature similarity is less than 30%, the correction factor does not exceed 0.3.

[0039] If there is a spatial occlusion relationship between targets in adjacent grids, the correction factor is adjusted according to the type of occlusion: if the occlusion is vegetation, the expression for the correction factor is: correction factor = 0.4 - 0.1 × vegetation coverage; if the occlusion is terrain cutting, the expression for the correction factor is: correction factor = 0.2 - 0.1 × terrain drop ratio.

[0040] Preferably, the method of correcting the confidence level of the recognition result output by the target recognition model using the correlation confidence level includes the following steps:

[0041] Preset association threshold parameters based on geological scene type;

[0042] When the association confidence is not lower than the association threshold parameter, the original output confidence of the target in the adjacent grid is positively corrected. The expression of the corrected confidence is: Corrected confidence = Original confidence × (1 + Association confidence × Association type coefficient), where the association type coefficient is set according to the geological scene type.

[0043] When the association confidence is lower than the association threshold parameter, the original output confidence of the target in the adjacent grid is reversed. The expression for the corrected confidence is: Corrected confidence = Original confidence × (1 - (Association threshold parameter - Association confidence) × 0.5);

[0044] If there are multiple sets of target associations in adjacent grids, the weighted average of the confidence scores of each set of associations is taken as the comprehensive association factor, and the original output confidence scores of the targets in each set of adjacent grids are calibrated a second time.

[0045] Beneficial Effects: The model-complementary target recognition method of this application, by standardizing the input image, provides a more consistent image input for subsequent target recognition, reducing feature extraction bias caused by differences in image size and color value fluctuations, and mitigating the impact of unstable image input on recognition results in complex geological scenes. Secondly, by using a target recognition model to perform grid division and feature extraction on the standardized image, the method identifies the existence, category, and location bounding box of landslide signs within each grid and outputs confidence scores. Compared to traditional fixed grid division and single feature extraction modes, this method can more systematically capture landslide features, providing a more accurate preliminary identification basis for subsequent screening and helping to reduce the risk of missing small landslide features. Simultaneously, by retaining valid preliminary results based on a set confidence threshold, it can selectively filter out low-confidence misjudgments, reducing... It reduces identification bias caused by invalid information interference, is more flexible than fixed threshold screening, and can better adapt to the screening needs of different confidence features in landslide detection. In addition, when the effective preliminary results are empty or there are low-confidence candidate targets, a large model with a pre-configured scene knowledge base is called for secondary detection. The large model uses multi-dimensional verification rules based on scene type to jointly analyze the global features and local details of the image, which can supplement the recognition ability of traditional models in complex geological scenes and reduce missed detections or misjudgments caused by the limitations of a single model. Furthermore, the fusion of effective preliminary results and large model detection results based on the weight allocation mechanism can comprehensively leverage the preliminary recognition advantages of the target recognition model and the scene complementarity advantages of the large model, outputting comprehensive recognition results that are more in line with the actual geological scene, effectively improving the overall accuracy and robustness of landslide detection. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a flowchart illustrating a method for identifying targets based on model complementarity, as provided in an embodiment of this application. Detailed Implementation

[0048] The technical solutions in the embodiments of this application will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0049] In this document, the term "comprising" is intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0050] This embodiment discloses, in a first aspect, a method as follows: Figure 1 The method for target identification based on model complementarity, as shown, achieves high-precision target identification in complex scenes through the complementary collaboration between traditional target identification models and large-scale models, making it particularly suitable for geological disaster detection scenarios such as landslides. Its core process includes: image preprocessing, preliminary target identification by the target identification model, confidence filtering, large-scale model scene complementarity detection, and multi-model result fusion, ultimately outputting a comprehensive identification result.

[0051] In detail

[0052] The image preprocessing steps specifically involve: standardizing the input image, which includes adjusting the input image to a preset uniform size and normalizing the color values. Specifically, the preset uniform size involves scaling the input image to a preset size (configurable as 640x640 pixels, 512x512 pixels, or 800x800 pixels; 640x640 is preferred for landslide detection to balance accuracy and speed). Color value normalization involves converting the RGB pixel values ​​from 0-255 into a standardized range of 0-1.

[0053] In addition, while performing standardization, adaptive enhancement operations are performed to optimize image features in complex scenes;

[0054] The adaptive enhancement operations include at least one of the following: adaptive denoising based on noise detection (such as removing salt-and-pepper noise or motion blur in the image through Gaussian filtering or nonlocal mean denoising algorithms), contrast enhancement for low-light scenes (such as using the Retinex algorithm to improve contrast and highlight key features such as surface cracks in images taken on rainy days), edge sharpening for complex backgrounds (such as using the Laplacian operator to enhance the boundary contours between landslide bodies and normal mountains), and local feature magnification for small targets (such as locally magnifying small areas (<32x32 pixels) in the image that appear to be piles of rubble through super-resolution reconstruction technology to improve subsequent recognition accuracy).

[0055] The initial identification steps of the target recognition model are as follows: the target recognition model divides the standardized image into grids and extracts features, identifies the existence, category, and location box of the target within each grid, and outputs the confidence level of the corresponding recognition result.

[0056] The process of dividing the standardized image into grids and extracting features includes the following steps:

[0057] Based on the target density and size distribution characteristics of the standardized image, the grid size and density are adjusted, including: areas with small targets (such as loose gravel) are divided into high-density fine-grained grids (such as 16x16 pixels), and areas without targets or with large targets are divided into low-density coarse-grained grids (such as 64x64 pixels).

[0058] By fusing shallow, low-dimensional edge features (such as the contour lines of a landslide) with deep, high-dimensional semantic features (such as the abstract features of a landslide), the feature weights of key regions of the target (such as the target contour and feature points) are enhanced through an attention mechanism to improve the recognition accuracy of small targets and partially occluded targets in complex scenes and complete feature extraction.

[0059] In addition, when identifying the existence, category, and location bounding box of targets within each grid, the target recognition model also uses the constructed inter-grid feature correlation matrix to perform correlation reasoning on the target features of adjacent grids.

[0060] Specifically, associative reasoning includes:

[0061] Determine the association type, which includes whether there is a spatial occlusion relationship between targets in adjacent grids (e.g., whether "vegetation area occludes landslide traces"), whether they belong to different parts of the same target (e.g., whether "exposed soil in adjacent grids belongs to the same landslide body"), and whether they have a semantic logical association.

[0062] Based on the judgment result of the association type, the judgment of the target category and the coordinates of the location box are corrected. At the same time, the association confidence is calculated, and the confidence of the recognition result output by the target recognition model is corrected by the association confidence to improve the recognition ability of target integrity and scene association. At this time, the confidence of the recognition result output by the target recognition model is the corrected result. The correction of the recognition result is as follows: if two grids are determined to be different parts of the same landslide, their location boxes are merged and the category is corrected to "landslide". If there is occlusion, the type of occlusion is marked (such as "shrub occlusion").

[0063] A feasible approach is to determine the association confidence level by including the following steps:

[0064] Extract multi-dimensional features of targets in adjacent grids. These multi-dimensional features include spatial location features (such as grid center distance and boundary overlap), visual features (such as color distribution similarity and texture consistency), and semantic features (such as the common association probability of targets in a preset geological scene).

[0065] Construct a feature association scoring matrix, assign feature weights to the extracted multi-dimensional features (the weights of spatial location features are adaptively adjusted according to the target size, and the weights of semantic features are updated in real time based on the geological scene knowledge base), and calculate the initial association score by weighted summation;

[0066] Based on the association type, a correction factor is set to calibrate the initial association score, resulting in the final association confidence level (range 0-1), where:

[0067] If targets within adjacent grids are determined to belong to different parts of the same target, the correction factor is dynamically adjusted based on the similarity of geological features (such as lithological consistency and displacement direction correlation): when the feature similarity is not less than 80%, the correction factor is not less than 0.7; when the feature similarity is between 30% and 80%, the expression for the correction factor is: correction factor = 0.3 + 0.8 × (feature similarity - 0.3); when the feature similarity is less than 30%, the correction factor does not exceed 0.3.

[0068] If there is a spatial occlusion relationship between targets in adjacent grids, the correction factor is adjusted according to the type of occlusion: if the occlusion is vegetation (such as shrubs or forests), (the higher the coverage, the lower the factor), the expression for the correction factor is: correction factor = 0.4 - 0.1 × vegetation coverage; if the occlusion is terrain cutting (such as valleys or cliffs), the expression for the correction factor is: correction factor = 0.2 - 0.1 × terrain drop ratio (the greater the drop, the lower the factor).

[0069] Furthermore, the confidence level of the target recognition model's output is corrected by associating the confidence level with the target recognition model, including the following steps:

[0070] Preset association threshold parameters according to the geological scene type (e.g., for scenes with dense targets and complex relationships (such as urban traffic intersection monitoring), the preset association threshold parameter is 0.4 to 0.5 to balance the sensitivity and accuracy of association judgment; for scenes with sparse targets and strict control of false associations (such as precision inspection of industrial equipment), the preset association threshold parameter is 0.6 to 0.7 to prioritize the accuracy of association judgment; the preset fixed values ​​are determined based on the statistical average of historical association samples for the corresponding scene, and can be manually fine-tuned through the manual configuration interface to adapt to the special needs of the scene).

[0071] When the association confidence level is not lower than the association threshold parameter, the original output confidence level of the target in the adjacent grid is positively corrected. The expression for the corrected confidence level is: Corrected confidence level = Original confidence level × (1 + Association confidence level × Association type coefficient), where the association type coefficient is set according to the geological scene type (e.g., if the association type is 'different parts of the same landslide body', the coefficient = 1.2; if the association type is 'geological genetic chain association' (e.g., causal relationship of crack-slip zone-deposit), the coefficient = 1.5; if the association type is 'regional tectonic association' (e.g., different landslide bodies share the same fault zone), the coefficient = 1.3; if the association type is 'spatial occlusion relationship', the coefficient = 0.8).

[0072] When the association confidence is lower than the association threshold parameter, the original output confidence of the target in the adjacent grid is reversed. The expression for the corrected confidence is: Corrected confidence = Original confidence × (1 - (Association threshold parameter - Association confidence) × 0.5), and the corrected confidence is not lower than 0.1.

[0073] If there are multiple sets of target associations in adjacent grids, the weighted average of the association confidence scores of each set is taken as the comprehensive association factor. The original output confidence scores of targets in each set of adjacent grids are then calibrated twice to eliminate the influence of isolated grid features on the misjudgment of confidence scores.

[0074] The confidence filtering step specifically involves retaining identification results with confidence levels exceeding a set confidence threshold as valid preliminary results, based on a pre-defined confidence threshold. The confidence threshold is adjusted according to the target category, including: for preset high-priority targets (such as landslide areas or hazardous objects), a threshold no lower than the high-priority target threshold parameter (e.g., 0.6) is used; for regular targets, a threshold no lower than the regular target threshold parameter (e.g., 0.5) is used. Furthermore, the threshold can be iteratively updated using model training data and validation results. For example, after each batch of detections, the threshold is automatically lowered by 0.05 based on the validation set accuracy (e.g., when the landslide identification false negative rate is >5%), and vice versa, by 0.03.

[0075] The large-scale model scene complementary detection step is as follows: when the effective preliminary result output by the confidence filtering step is empty (e.g., the target recognition model did not identify the landslide) or there are low-confidence candidate targets (e.g., a suspected area with a confidence of 0.45), a large model (such as any multimodal model in the prior art) with a pre-configured geological scene knowledge base is called for secondary detection. The secondary detection includes the large model calling the corresponding multi-dimensional verification rules based on the geological scene type to jointly analyze the global features and local details of the image.

[0076] The large model includes preset geological scenarios and their corresponding validation rule base. The validation rule base must include geological professional knowledge dimensions based on expert experience: a stratigraphic lithology database (such as landslide susceptibility parameters for loose deposits and hard rock masses), a rainfall-landslide correlation model (such as the quantitative relationship between hourly rainfall intensity and soil saturation), and a historical landslide case database (including correlation data on landslide types, triggering factors, and morphological characteristics).

[0077] Furthermore, the multi-dimensional verification rules include:

[0078] Scene feature verification: Identify the distribution characteristics and proportion of large areas of exposed soil or gravel (e.g., "the exposed area accounts for 35%, which meets the characteristics of a landslide"); detect whether the terrain has unnatural tilt, fracture marks or stepped deformation (e.g., if the slope is >30° and there are fracture lines, it is marked as high risk through the slope calculation model).

[0079] Spatiotemporal comparison verification: Compare the current image with historical images to analyze changes in vegetation cover and soil displacement trajectory (if the vegetation reduction rate is >40%, it will indicate "potential landslide precursors");

[0080] Semantic reasoning verification: Based on geological logic reasoning, the morphological characteristics of potential landslide risk areas are inferred through regional geological data (such as "this area is a fault zone with frequent historical landslides").

[0081] The multi-model result fusion step is as follows: the effective preliminary results and the detection results of the large model are fused based on the weight allocation mechanism, and a comprehensive recognition result is output; among which, the weights are adjusted in real time according to the target category, scene complexity and historical model performance.

[0082] The weight allocation mechanism includes the basis for weight allocation, the rules for weight adjustment, and the strategies for weight optimization.

[0083] The weight allocation is configured based on target category, scene complexity, and the model's historical recognition accuracy. Target category includes known target categories trained on the target recognition model (e.g., "landslide"), and untrained target categories identified by the large model (e.g., "frozen soil landslide"). For known categories, the target recognition model's results are prioritized (60% weight), while new categories are prioritized (50% weight). Scene complexity includes whether the target is occluded, whether it is a small target, and the degree of background interference. If scene complexity exists, the large model's weight is increased to 55%. The model's historical recognition accuracy is the statistical value of the target recognition model and the large model's historical correct recognition rates for similar targets or scenes. If the target recognition model's accuracy is 90% and the large model's is 85% in similar scenes, the target recognition model's weight is increased by 5%.

[0084] The weight adjustment rules include: When the target recognition model's confidence level for a known category of target is not lower than the high confidence parameter for the known category (e.g., 0.8), and its historical recognition accuracy is not lower than a preset benchmark (e.g., 90%), the weight percentage of the effective preliminary results output by the target recognition model is not lower than the high weight parameter of the target recognition model (i.e., 60%). When the semantic inference confidence level of the large model for a new category of target or a target in a complex scene (with occlusion, small targets, or high background interference) is not lower than the high confidence parameter for the new category (e.g., 0.7), and its historical recognition accuracy is not lower than a preset benchmark (e.g., 85%), the weight percentage of the detection results output by the large model is not lower than the high weight parameter of the large model (i.e., 50%). Furthermore, when there is a conflict between the output results of the target recognition model and the large model (e.g., the target recognition model identifies it as a "normal mountain," while the large model identifies it as a "landslide"), the weight is tilted towards the model with the higher historical recognition accuracy (e.g., if the large model has a higher accuracy, its weight percentage becomes 60%).

[0085] Among these, the high-weight parameter of the target recognition model is the minimum weight percentage threshold set for the "effective preliminary results" output by the target recognition model. When the confidence level of the target recognition model in recognizing known category targets reaches the "known category high confidence parameter," and its historical recognition accuracy exceeds the "preset benchmark value," it indicates that the model performs reliably in recognizing the current known category targets (consistent with its advantage in quickly detecting known categories). At this time, the weight percentage of its output results is limited by the "target recognition model high weight parameter" to ensure the stability and dominance of the known category target recognition results (for example, preset to 60%, meaning that the weight of the target recognition model results accounts for at least 60%). The high-weight parameter of the large model is the minimum weight percentage threshold set for the "detection results" output by the large model. When the semantic reasoning confidence of the large model for new categories of targets or targets in complex scenes (such as those with occlusion or small targets) reaches the "high confidence parameter for new categories," and its historical recognition accuracy exceeds the "preset benchmark value," it indicates that the large model's advantage in handling new categories or complex scenes has been effectively demonstrated. At this point, the weight of its output result is limited to a certain percentage by the "high weight parameter for large models," ensuring that the recognition results of the large model in its advantageous scenarios can fully influence the final result (e.g., preset to 50%, meaning that the weight of the large model's result is at least 50%). Furthermore, the specific value of the "high weight parameter" is configurable and can be adjusted according to the actual application scenario (such as landslide detection), the importance of the target category, and the model's historical performance (e.g., through reinforcement learning algorithms based on error feedback iterative optimization) to balance the complementarity of the two types of models in different scenarios, ultimately improving the accuracy and robustness of the comprehensive recognition.

[0086] Weight optimization strategies include: based on reinforcement learning algorithms (such as DQN-Deep Q Network), using the error between the comprehensive recognition result and the manually labeled result as the reward signal, periodically updating the parameters in the weight adjustment rules (including high confidence parameters of known categories, high weight parameters of target recognition models, etc.) to make the weight allocation adapt to scene changes.

[0087] Using the above-mentioned model-complementary target identification method, taking landslide detection in a mountainous area as an example, the implementation process is as follows:

[0088] The input image was captured by the drone, preprocessed to 640x640 pixels, and normalized to enhance the contrast in low-light areas.

[0089] The target recognition model initially identified a "suspected landslide body (confidence level 0.58)", which was below the high priority threshold (0.6), triggering large model detection.

[0090] The large model outputs "high risk of landslide (confidence 0.82)" through scene feature verification (40% of the area is bare), spatiotemporal comparison (38% reduction in vegetation) and semantic reasoning (fault zone area).

[0091] Fusion weights: target recognition model weight 40%, large model weight 60%, final output "This area is a landslide, overall confidence level 0.72".

[0092] In summary, the target recognition method based on model complementarity in this embodiment not only leverages the real-time advantages of traditional target recognition models, but also utilizes the global analysis and reasoning capabilities of large models to compensate for the shortcomings in recognition under complex scenarios, significantly improving the accuracy and robustness of target recognition.

[0093] In the embodiments provided in this application, it should be understood that the embodiments described herein can be implemented in hardware, software, firmware, middleware, code, or any suitable combination thereof. For hardware implementation, the processor may be implemented in one or more of the following: application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), processors, controllers, microcontrollers, microprocessors, other electronic units designed to implement the functions described herein, or combinations thereof. For software implementation, some or all of the processes of the embodiments may be performed by a computer program instructing the associated hardware. During implementation, the program may be stored in a computer-readable storage medium or transmitted as one or more instructions or code on a computer-readable storage medium. Computer-readable storage media include computer storage media and communication media, wherein communication media include any medium that facilitates the transmission of a computer program from one place to another. Storage media may be any available medium accessible to a computer. Computer-readable storage media may include, but are not limited to, RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code having the form of instructions or data structures and accessible to a computer.

[0094] Finally, it should be noted that the above description is only a preferred embodiment of this application and is not intended to limit this application. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for identifying targets based on model complementarity, characterized in that, The method includes the following steps: Image preprocessing steps: Standardize the input image, which includes adjusting the input image to a preset uniform size and normalizing the color values; Preliminary identification steps of the target recognition model: The target recognition model divides the standardized image into grids and extracts features, identifies the existence, category, and location box of the target within each grid, and outputs the confidence level of the corresponding recognition result; Confidence filtering step: Based on the set confidence threshold, retain the recognition results with a confidence level exceeding the threshold as valid preliminary results; Large model scene complementarity detection step: When the valid preliminary result output by the confidence filtering step is empty or there are low-confidence candidate targets, the large model with the pre-configured geological scene knowledge base is called to perform secondary detection. The secondary detection includes the large model calling the corresponding multi-dimensional verification rules based on the geological scene type to jointly analyze the global features and local details of the image. Multi-model result fusion step: Based on the weight allocation mechanism, the effective preliminary results are fused with the detection results of the large model, and a comprehensive recognition result is output; wherein, the weights are adjusted in real time according to the target category, scene complexity and historical model performance.

2. The method for identifying targets based on model complementarity according to claim 1, characterized in that, The image preprocessing step further includes: performing adaptive enhancement operations to optimize image features in complex scenes while performing standardization processing; The adaptive enhancement operation includes: adaptive denoising based on noise detection, and at least one of the following: contrast enhancement for low-light scenes, edge sharpening for complex backgrounds, and local feature magnification for small targets.

3. The method for identifying targets based on model complementarity according to claim 1, characterized in that, In the confidence filtering step, the confidence threshold is adjusted according to the target category, including: for preset high-priority targets, a threshold no lower than the high-priority target threshold parameter is used; for regular targets, a threshold no lower than the regular target threshold parameter is used; and the threshold can be iteratively updated through model training data and validation results.

4. The method for identifying targets based on model complementarity according to claim 1, characterized in that, In the large-model scene complementarity detection step, the large model includes a preset geological scene and its corresponding verification rule library; the multi-dimensional verification rules include: Scene feature verification: Identify the distribution characteristics and proportion of large areas of exposed soil or gravel; detect whether the terrain has unnatural tilts, cracks, or stepped deformations; Spatiotemporal comparison verification: Compare the current image with historical images to analyze changes in vegetation cover and soil displacement trajectory; Semantic reasoning verification: Based on geological logic reasoning, inferring the morphological characteristics of potential landslide risk areas through regional geological data.

5. The method for identifying targets based on model complementarity according to claim 1, characterized in that, The weight allocation mechanism includes the basis for weight allocation, weight adjustment rules, and weight optimization strategies; The weight allocation is configured as follows: weights are determined based on target category, scene complexity, and model historical recognition accuracy; wherein, the target category includes known target categories that the target recognition model has trained on, and untrained target categories that the large model recognizes; the scene complexity includes whether the target is occluded, whether it is a small target, and the degree of background interference; the model historical recognition accuracy is the statistical value of the historical correct recognition rate of the target recognition model and the large model in the same type of target or the same scene; The weight adjustment rules include: when the confidence level of the target recognition model for a known category of target is not lower than the high confidence parameter of the known category, and its historical recognition accuracy is not lower than the preset benchmark value, then the weight ratio of the effective preliminary results output by the target recognition model is not lower than the high weight parameter of the target recognition model; when the semantic reasoning confidence level of the large model for a new category of target or a target in a complex scene is not lower than the high confidence parameter of the new category, and its historical recognition accuracy is not lower than the preset benchmark value, then the weight ratio of the detection results output by the large model is not lower than the high weight parameter of the large model; when there is a conflict between the output results of the target recognition model and the large model, the weight is tilted towards the model with higher historical recognition accuracy. The weight optimization strategy includes: based on a reinforcement learning algorithm, using the error between the comprehensive recognition result and the manually labeled result as the reward signal, and periodically updating the parameters in the weight adjustment rule.

6. The method for identifying targets based on model complementarity according to claim 1, characterized in that, In the preliminary recognition step of the target recognition model, the standardized image is divided into grids and its features are extracted, including the following steps: The grid size and density are adjusted based on the target density and size distribution characteristics of the standardized image, including: dividing regions with small targets into high-density fine-grained grids, and dividing regions without targets or with large targets into low-density coarse-grained grids. Based on the fusion of shallow low-dimensional edge features and deep high-dimensional semantic features, the feature weights of key regions of the target are enhanced through an attention mechanism to complete feature extraction.

7. The method for identifying targets based on model complementarity according to claim 1 or 6, characterized in that, In the preliminary identification step of the target recognition model, when identifying the existence, category, and location box of the target in each grid, the target recognition model also performs correlation reasoning on the target features of adjacent grids through the constructed inter-grid feature correlation matrix.

8. The method for identifying targets based on model complementarity according to claim 7, characterized in that, The related reasoning includes: Determine the association type, which includes whether there is a spatial occlusion relationship between targets in adjacent grids, whether they belong to different parts of the same target, and whether they have a semantic logical relationship. Based on the judgment result of the association type, the judgment of the target category and the coordinates of the location box are corrected. At the same time, the association confidence is calculated, and the confidence of the recognition result output by the target recognition model is corrected by the association confidence.

9. The method for identifying targets based on model complementarity according to claim 8, characterized in that, Determining the association confidence level includes the following steps: Extract multi-dimensional features of targets in adjacent grids, including spatial location features, visual features, and semantic features; Construct a feature association scoring matrix, assign feature weights to the extracted multi-dimensional features, and calculate the initial association score by weighted summation; A correction factor is set based on the association type to calibrate the initial association score, resulting in the final association confidence score, where: If targets within adjacent grids are determined to belong to different parts of the same target, the correction factor is dynamically adjusted based on the similarity of geological features: when the feature similarity is not less than 80%, the correction factor is not less than 0.7; when the feature similarity is between 30% and 80%, the expression for the correction factor is: correction factor = 0.3 + 0.8 × (feature similarity - 0.3); when the feature similarity is less than 30%, the correction factor does not exceed 0.

3. If there is a spatial occlusion relationship between targets in adjacent grids, the correction factor is adjusted according to the type of occlusion: if the occlusion is vegetation, the expression for the correction factor is: correction factor = 0.4 - 0.1 × vegetation coverage; if the occlusion is terrain cutting, the expression for the correction factor is: correction factor = 0.2 - 0.1 × terrain drop ratio.

10. The method for identifying targets based on model complementarity according to claim 8, characterized in that, The method of correcting the confidence level of the target recognition model output by means of the association confidence level includes the following steps: Preset association threshold parameters based on geological scene type; When the association confidence is not lower than the association threshold parameter, the original output confidence of the target in the adjacent grid is positively corrected. The expression of the corrected confidence is: Corrected confidence = Original confidence × (1 + Association confidence × Association type coefficient), where the association type coefficient is set according to the geological scene type. When the association confidence is lower than the association threshold parameter, the original output confidence of the target in the adjacent grid is reversed. The expression for the corrected confidence is: Corrected confidence = Original confidence × (1 - (Association threshold parameter - Association confidence) × 0.5); If there are multiple sets of target associations in adjacent grids, the weighted average of the confidence scores of each set of associations is taken as the comprehensive association factor, and the original output confidence scores of the targets in each set of adjacent grids are calibrated a second time.

Images