A target detection robustness augmented sample selection method based on two-dimensional performance surface

Abstract

The application is a target detection robustness augmented sample selection method based on a two-dimensional performance surface. The application relates to the technical field of computer vision and deep learning, and the application constructs a robustness augmented sample candidate pool, filters out candidate samples with the most value for improving the robustness of a model, performs genetic algorithm-based augmented sample optimization selection, outputs an augmented sample set corresponding to a chromosome with the highest fitness, performs basic training set reduction based on active learning, filters out a representative sample subset from an original training set to construct a basic training set, performs model robustness enhancement training, constructs a robustness enhancement training set, and trains a target detection model. The application overcomes the problem of insufficient robustness of a target detection model under multi-dimensional complex interference conditions.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and deep learning technology, and is a robust augmented sample selection method for target detection based on two-dimensional performance surfaces. Background Technology

[0002] Object detection is one of the core tasks in computer vision, aiming to locate and identify objects of interest from images. In recent years, deep learning object detection methods, represented by the YOLO series and Faster R-CNN, have made significant progress on multiple benchmark datasets. However, these models often face the challenge of image quality degradation in real-world deployments, exhibiting vulnerability to input perturbations.

[0003] Existing research on improving model robustness mainly focuses on two directions: one is to enhance robustness by optimizing the model structure, and the other is to improve the model's resistance to interference by constructing robustness-enhancing datasets. The former usually requires deep modification of the network structure for specific detection tasks, which is complex and has poor transferability; the latter introduces diverse perturbation samples through data augmentation, which is relatively easier to implement and therefore widely adopted.

[0004] However, existing technical solutions have the following significant drawbacks: First, at the robustness assessment level, existing methods primarily perform performance analysis under a single disturbance dimension. Traditional robustness assessments typically examine the independent impact of only one type of disturbance (such as noise, blur, or illumination changes) on model performance each time, in which case the model performance can be represented by a single curve. However, in real-world applications, target detection models are often simultaneously affected by the coupled effects of multiple disturbance factors. Especially in remote sensing image target detection tasks, models need to handle multiple disturbance sources such as atmospheric conditions, sensor noise, and geometric distortion. Studies have shown that the performance degradation of models under multi-dimensional disturbances is not a simple linear superposition of two single factors, but exhibits a significant nonlinear coupling effect. That is, when an image is simultaneously subjected to high-intensity noise interference and extreme brightness changes, the detection accuracy will drop precipitously, and the damage to the model's feature extraction ability from multiple sources has a multiplicative effect. Existing methods struggle to effectively characterize the model's performance response characteristics in a multi-dimensional disturbance parameter space.

[0005] Second, at the sample selection level, existing data augmentation methods lack targeted selection mechanisms for regions with weak model robustness. Traditional data augmentation strategies include basic operations such as random cropping, flipping, and color transformation, as well as hybrid augmentation methods such as CutMix, Mixup, and AugMix. These methods generate augmented samples using random or heuristic rules, failing to identify and prioritize key samples that can effectively improve the model's resistance to interference. In practical applications, the number of augmented samples cannot be expanded indefinitely; how to effectively improve model robustness under a limited number of augmented samples becomes a key challenge. Although active learning and Coreset selection methods have made significant progress in optimizing training efficiency, these methods mainly focus on general sample value metrics and lack correlation analysis with regions with weak model robustness.

[0006] Third, at the level of augmented sample combination optimization, existing methods lack a systematic global optimization strategy. When the candidate sample pool is large, the feasible solution space exhibits combinatorial explosive growth, and selecting the optimal augmented sample subset from a large number of candidate samples is an NP-hard problem. Existing methods mainly rely on greedy algorithms or simple random sampling strategies, which are prone to getting trapped in local optima and cannot guarantee the optimality of the selected augmented sample combination in improving the overall robustness of the model.

[0007] In summary, existing technical solutions have significant shortcomings in robustness assessment under complex interference conditions, efficient sample selection, and global optimization of augmentation strategies, making it difficult to effectively address the model performance degradation caused by multidimensional complex interference in real-world application scenarios. Summary of the Invention

[0008] This invention addresses the problem of insufficient robustness of target detection models under multidimensional compound interference conditions. It identifies weak areas of robustness in the model through performance surface analysis, constructs a high-value candidate sample pool, and uses a genetic algorithm to optimize and select the augmented sample subset that best improves the robustness of the model. This invention discloses a target detection robustness augmented sample selection method based on two-dimensional performance surface.

[0009] This invention provides the following technical solutions: A robust augmented sample selection method for target detection based on two-dimensional performance surfaces, the method comprising the following steps: Step 1: Construct a robust augmentation sample candidate pool and select the candidate samples that are most valuable for improving the robustness of the model; Step 2: Optimize the selection of augmented samples based on genetic algorithm, and output the augmented sample set corresponding to the chromosome with the highest fitness; Step 3: Based on active learning, reduce the basic training set by selecting a representative subset of samples from the original training set to construct the basic training set; Step 4: Perform robustness enhancement training on the model, construct a robustness enhancement training set, and train the object detection model.

[0010] Preferably, let For target detection models, For the test set, and There are two types of interference, and their intensity parameters are as follows: and The two-dimensional performance response function is defined as follows:

[0011] in For two-dimensional interference parameter space, To the interference parameters The generated interference test set; Performance response function A two-dimensional surface is formed in the parameter space, and its geometric shape reflects the model's response characteristics to different combinations of disturbances; Four representative types of interference in remote sensing imaging were selected, namely: , , , Based on these four types of interference, four sets of composite interference scenarios are constructed by combining them in pairs:

[0012] Let the performance response function be defined. At point If the model is continuously differentiable at a point, its local robustness at that point is defined as follows:

[0013] in, It is a unit direction vector. This is a preset performance tolerance threshold; The measurement of performance degradation exceeding [a certain value] in the most unfavorable direction was taken. Minimum required disturbance amount; Let the performance surface There are a total of Given a set of grid points, sort all grid points in descending order of sensitivity. Let the sorted sensitivity sequence be denoted as . ,in The cumulative sensitivity contribution rate function is defined as:

[0014] The function describes the order of sorting before and after sorting. The contribution percentage of each highly sensitive point to the total sensitivity, given a target contribution rate. Number of key sampling points Defined as the cumulative contribution rate reaching a certain level for the first time. Minimum number of points:

[0015] The corresponding adaptive sensitivity threshold is ,parameter It controls the trade-off between the size and coverage of the candidate sample pool; For performance surfaces The set of key sampling points is defined as follows:

[0016] Each key sampling point Corresponding to a set of samples under specific combined interference conditions Its size is equal to the number of images in the original test set. ; For a single performance surface, its candidate sample set is:

[0017] The candidate sample pool is formed by integrating four sets of composite interference scenarios:

[0018] Assume that the average selection for each performance surface is... If there are 1 key sampling point, then the size of the candidate sample pool satisfies:

[0019] The equality holds when there is no overlap at the key sampling points of each surface.

[0020] Preferably, the candidate sample pool is set as follows: The budget for expanding the sample is The basic training set is The problem of augmenting the sample selection is formalized as follows:

[0021] Wherein, fitness function Defined as using augmented dataset Detection performance of the trained model on the validation set:

[0022] The size of the feasible solution space of the problem is ;when , In such cases, heuristic optimization strategies should be employed.

[0023] Preferably, the constructed genetic algorithm uses fixed-length integer encoding, with one chromosome... Indicated as containing The ordered sequence of one gene:

[0024] Each gene Indicates the index of the selected sample in the candidate pool; chromosome The corresponding augmented sample set is To avoid repeated selection of the same sample, it is required that the genes within the chromosome are distinct, i.e. For all Established; Let the sample The type of composite interference is Interference type The frequency in the candidate pool is defined as:

[0025] sample The selection weight is defined as:

[0026] in To act as a smoothing factor and avoid numerical problems in the zero-frequency case; the normalized selection probability is:

[0027] in This is the current set of available samples; chromosome The fitness value is defined as the detection performance of the augmented model on the validation set:

[0028] in, For training on mixed training sets The model after the wheel, The mean precision is the average value of the IoU threshold ranging from 0.50 to 0.95. Let the training time be... Population size is The maximum number of generations is The total evaluation time is:

[0029] A tournament selection strategy is used, with random selection from the population each time. Individuals form a competitive set, and the one with the highest fitness is selected as the parent:

[0030] Suppose that individuals in the population are arranged in descending order of fitness, and the rank is... The probability that an individual is selected is:

[0031] Using a single-point crossover strategy, given two parent chromosomes... and In the interval Randomly select intersection points Gene fragments exchanged at crossover points generate offspring:

[0032]

[0033] Crossover operations with probability The core function of execution and crossover operations is to combine the superior gene patterns of different parents to achieve efficient exploration of the solution space; A point mutation strategy is employed, where each gene locus in the chromosome is independently assessed for mutation, using probability. Replace it with a random index from the candidate pool:

[0034] in, These are independent random numbers; Preferably, a deterministic repair mechanism is designed: Let chromosomes There are duplicate gene sets in it. For each repeating position ,Will Replace with random samples that were not selected from the candidate pool:

[0035] Repair mechanisms ensure that chromosomes always meet the constraints after genetic manipulation.

[0036] Preferably, in each generation of evolution, the top-ranked fitness values ​​are... Individuals are directly replicated into the next generation of the population. Let be the elite retention rate. Under an elite retention strategy, the optimal fitness value of the population monotonically does not decrease with each generation, i.e.:

[0037] in ; Let the first The optimal individual in the generation is The strategy of retaining elites It was directly copied to the first page. In the population; therefore:

[0038] Elite retention rate With the rate of variation Together, they determine the diversity-convergence tradeoff; let... For the first Regarding population diversity, then:

[0039] when When it is too big, When the population tends to be homogeneous, the algorithm is prone to getting trapped in local optima; when When the size is too large, the optimal gene pattern is frequently disrupted, and the convergence efficiency decreases. Termination condition: The algorithm terminates when any of the following conditions are met: (1) Upper limit of generation: The number of generations of evolution reaches the preset maximum value. ; (2) Convergence criterion: Continuous The improvement in optimal fitness within a generation is less than the threshold. :

[0040] The convergence criterion is based on the following observation: when the population is close to the optimal solution, the rate of fitness improvement tends to level off, and the marginal benefit of further evolution diminishes; After the algorithm terminates, it outputs the optimal chromosome that appears during the evolutionary process and its corresponding augmented sample set:

[0041] Preferably, the population size of the genetic algorithm is set to be... The maximum number of generations is The original training set size is The training time for a single sample is The number of training rounds is The total computation time of the genetic algorithm is:

[0042] Let the original training set be ,in For image, For the corresponding object detection annotations, the training set reduction problem is formalized as follows:

[0043] in, For the target size, For the information loss function, the metric uses subsets. The trained model is relative to using the full set Performance gap between trained models:

[0044] in, Indicates in the dataset The model trained on it, To predict the measure of difference; Model For the sample The predicted distribution is ,sample The value of information can be measured by expected information gain:

[0045] in, EIG is the entropy function; it measures the observed samples. The expected reduction in model uncertainty after the true labels are obtained; Model For the sample The test results are ,in For bounding box, For classification confidence, To determine the number of targets to be detected, define a sample. The overall uncertainty score is:

[0046] in, These are the weighting coefficients. and These are the confidence level variance term and the low confidence level proportion term, respectively. When the model is in the sample No target was detected. This indicates that the sample poses an extreme challenge to the current model and should be assigned the highest level of uncertainty:

[0047] Uncertainty measures based on confidence level under the mild assumption. With expected information gain Positive correlation; Suppose the current basic training set Medium category The number of annotations is The total number of annotations is Then the category The annotation ratio is:

[0048] In each round of sample selection, a selection quota proportional to the current proportion is allocated to each category; let the target selection quantity for this round be... Then the category The quota is:

[0049] in, This is the floor function. Ensure each category receives at least one quota; reduce the dataset under the category balance constraint. The category distribution of the original dataset The category distributions are asymptotically consistent; Let the original dataset contain categories The proportion is During the iterative selection process, the category The cumulative number of selections is:

[0050] If the class distribution of the initial seed set is close ,but For all Established, thus .

[0051] Preferably, based on the detection performance of the pre-trained model, the samples are divided into three difficulty levels: Simple: All targets were detected with high confidence. ) Medium: Medium confidence level detection exists ( ) Challenges: Low-confidence detection or missed detections exist. or ) in, , The confidence threshold. This represents the actual target quantity. During the selection process within each category, stratified sampling is performed according to a preset difficulty distribution:

[0052] The iterative process terminates when either of the following conditions is met: the size of the reduced set reaches the target value. If the average uncertainty of newly added samples is lower than the preset threshold within several consecutive rounds, it indicates that the information is approaching saturation, and the marginal benefit of continuing to expand the training set is diminishing.

[0053] A computer-readable storage medium having a computer program stored thereon, the program being executed by a processor to implement a robust augmented sample selection method for target detection based on a two-dimensional performance surface.

[0054] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement a robust augmented sample selection method for target detection based on a two-dimensional performance surface.

[0055] The present invention has the following beneficial effects: This invention selects the most valuable candidate samples for improving model robustness from a complex interference space. First, four typical interference types are selected and combined in pairs to construct four sets of composite interference scenarios. Mesh sampling is performed on the two-dimensional parameter space of each composite interference set to evaluate the detection performance of the baseline model at each sampling point, thus constructing a two-dimensional performance surface. Subsequently, the interference sensitivity of each sampling point is calculated based on the gradient of the performance surface, and the cumulative sensitivity contribution rate criterion is used to adaptively select key sampling points. Finally, the interference samples corresponding to the key sampling points on multiple performance surfaces are aggregated to form a robust augmentation sample candidate pool. .

[0056] This invention uses candidate sample pools Selecting the optimal augmented sample subset is a large-scale combinatorial optimization problem. This paper uses a genetic algorithm to solve this problem: each sample selection scheme is encoded as a chromosome, and the detection performance (mAP) of the augmented model is used as the fitness function; a diverse initial population is generated through a strategy based on inverse frequency weighting; the sample combination is gradually optimized by iteratively performing genetic operations such as selection, crossover, and mutation; finally, the augmented sample set corresponding to the chromosome with the highest fitness is output. .

[0057] The iterative process of the genetic algorithm in this invention requires fitness evaluation of a large number of chromosomes, and each evaluation necessitates training a surrogate model. Using the complete original training set would incur prohibitive computational overhead. Therefore, this paper employs an active learning-based dataset reduction strategy: uncertainty metrics are used to identify samples with high informational value for model training; combined with class balance constraints and a difficulty-aware hierarchical selection mechanism, a representative subset of samples is selected from the original training set to construct the basic training set. This strategy significantly reduces the computational overhead of the genetic algorithm iteration process while maintaining model performance.

[0058] This invention utilizes the optimal augmented sample set output by the genetic algorithm. By merging the original training set with the target detection model's original training set, a robust enhanced training set is constructed, and the enhanced model is then trained. The enhanced model exhibits stronger performance stability under multi-dimensional compound disturbance conditions. Attached Figure Description

[0059] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0060] Figure 1 This is shown as the overall framework diagram of the present invention; Figure 2 This invention illustrates the method for constructing an augmented sample candidate pool based on performance surface sensitivity analysis. Figure 3 The diagram shows the augmented sample optimization selection method based on genetic algorithms. Figure 4 The diagram shows a framework for a basic training set reduction method based on active learning. Detailed Implementation

[0061] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] The present invention will be described in detail below with reference to specific embodiments. Specific Implementation Example 1: according to Figures 1 to 4 As shown, the specific optimized technical solution adopted by the present invention to solve the above-mentioned technical problems is: The present invention relates to a target detection robust augmented sample selection method based on two-dimensional performance surfaces.

[0064] This invention aims to identify weak regions in model robustness from a complex multidimensional disturbance space, providing a theoretical basis for targeted augmentation. The core idea is to construct a performance response surface under two-dimensional composite disturbance, use differential geometry tools to analyze the sensitivity of model performance to disturbance parameters, and then adaptively select the candidate samples most valuable for improving robustness. The overall framework is shown in Figure 2.

[0065] Two-dimensional composite interference performance surface modeling In practical applications, target detection models often face the coupled effects of multiple interference factors. Existing research mainly evaluates model performance under a single interference dimension, that is, fixing other factors and only varying the intensity of a certain interference, in which case the model performance can be represented as a one-dimensional response curve. However, this evaluation method ignores the interaction effects between interference factors. To more comprehensively characterize the performance degradation law of the model under multi-dimensional interference, this paper introduces the concept of a two-dimensional composite interference performance surface.

[0066] set up For target detection models, For the test set, and There are two types of interference, and their intensity parameters are as follows: and Define the two-dimensional performance response function as follows:

[0067] in For two-dimensional interference parameter space, To the interference parameters The generated interference test set. Performance response function A two-dimensional surface is formed in the parameter space, and its geometric shape reflects the model's response characteristics to different combinations of disturbances. Four representative types of interference in remote sensing imaging were selected, namely: , , , Based on these four types of interference, four sets of composite interference scenarios are constructed by combining them in pairs:

[0068] The design of this combination strategy follows the following considerations: (1) avoiding combinations of strongly correlated interference; (2) covering typical coupling modes such as "additive interference + multiplicative interference" and "frequency domain degradation + grayscale degradation".

[0069] Since direct analysis in a continuous parameter space is computationally infeasible, discretization sampling of the parameter space is necessary. This paper employs a uniform grid sampling strategy to... Discretize into Regular grid The choice of grid density involves a trade-off between sampling accuracy and computational cost.

[0070] Gradient-based interference sensitivity analysis The geometric features of performance surfaces contain rich information about model robustness. Intuitively, flat regions of the surface indicate that the model has a strong tolerance to the combination of disturbances, while steep regions reveal the model's vulnerabilities. This invention formalizes this intuitive understanding from the perspective of differential geometry, establishing a theoretical link between sensitivity measures and regions of weak robustness.

[0071] Let the performance response function be defined. At point It is continuously differentiable at that point. The local robustness of the model at that point is defined as:

[0072] in It is a unit direction vector. This is a preset performance tolerance threshold. Intuitively, The measurement of performance degradation exceeding [a certain value] in the most unfavorable direction was taken. The minimum amount of disturbance required. The smaller the value, the more susceptible the model is to disturbances, meaning the weaker its robustness. Proposition 1 (The Inverse Relationship Between Sensitivity and Local Robustness): Let... At point If it is second-order continuously differentiable, then the interference sensitivity is... With local robustness The following inverse relation is satisfied:

[0073] Therefore, high sensitivity ( Large) implies low local robustness ( Small), meaning that a highly sensitive area is equivalent to a weakly robust area. Proof: Assume sensitivity The steepest descent direction is:

[0074] along The direction is correct Expand Taylor:

[0075] Let the performance change The solution is that the performance degradation reaches [a certain value]. Critical perturbation:

[0076] By the definition of local robustness, Therefore:

[0077] This inequality shows that: sensitivity The larger the value, the better the local robustness. The smaller the upper bound, the better. hour, The model is unstable at this point for any small perturbation.

[0078] Inference 1: Sensitivity-based sample selection strategies prioritize coverage. The region with the smallest value is identified, thereby enabling targeted enhancement of weak links in robustness.

[0079] Adaptive identification of robust weak regions Based on the above sensitivity analysis, this invention proposes an adaptive weak region identification method. The core challenge lies in: how to determine the sensitivity threshold to distinguish between "weak regions" and "robust regions"? Fixed threshold strategies are difficult to adapt to the differences in characteristics of different performance surfaces. Therefore, this paper proposes an adaptive threshold determination method based on the cumulative sensitivity contribution rate.

[0080] The distribution of sensitivity on the performance surface typically exhibits a long-tail characteristic, meaning that a few highly sensitive regions contribute the majority of the overall sensitivity. This phenomenon can be explained by the Pareto Principle: the robustness bottleneck of the model tends to be concentrated in a limited number of critical regions, rather than being uniformly distributed throughout the entire disturbance space.

[0081] Let the performance surface There are a total of Given a set of grid points, sort all grid points in descending order of sensitivity. Let the sorted sensitivity sequence be denoted as . ,in The cumulative sensitivity contribution rate function is defined as:

[0082] This function describes the state of the sorted data. The contribution percentage of each highly sensitive point to the total sensitivity. Given a target contribution rate. Number of key sampling points Defined as the cumulative contribution rate reaching a certain level for the first time. Minimum number of points:

[0083] The corresponding adaptive sensitivity threshold is .parameter This controls the trade-off between the size and coverage of the candidate sample pool. A larger pool... The value will include more samples from moderately sensitive regions, improving coverage integrity but increasing the candidate pool size; smaller values... The value focuses on extremely weak areas, and the candidate pool is streamlined but may miss secondary weak links. For performance surfaces The set of key sampling points is defined as follows:

[0084] The sampling points in this set collectively contribute at least [amount] to the performance surface. The overall sensitivity of the proportion represents the core weak area of ​​the model under this complex interference scenario. Construction and Size Analysis of Candidate Sample Pool Based on the above analysis, the interference samples corresponding to the key sampling points identified on multiple performance surfaces are aggregated to construct the final robust augmented sample candidate pool.

[0085] The correspondence between sampling points and interference samples needs to be clearly defined: for each key sampling point Corresponding to a set of samples under specific combined interference conditions Its size is equal to the number of images in the original test set. . For a single performance surface, its candidate sample set is:

[0086] By integrating four sets of composite interference scenarios, the final candidate sample pool is:

[0087] Assume that the average selection for each performance surface is... If there are 1 key sampling point, then the size of the candidate sample pool satisfies:

[0088] The equality holds when there is no overlap at the key sampling points of the surfaces. This upper bound indicates that, through screening using the cumulative sensitivity contribution rate criterion, the candidate pool size increases from the original... Down to This effectively reduces the search space for subsequent genetic algorithms.

[0089] Genetic Algorithm-Based Augmented Sample Optimization Selection Based on the constructed candidate sample pool This invention investigates how to select the optimal augmented sample subset. The core challenge of this problem lies in the combinatorial explosion of the feasible solution space when the candidate pool is large, making exact solution computationally infeasible. This invention first formally models the problem and analyzes its computational complexity, then proposes a solution strategy based on a genetic algorithm, and finally theoretically analyzes the convergence of the algorithm and the quality assurance of the solution. The framework diagram is shown below. Figure 3 As shown.

[0090] 1. Modeling and Complexity Analysis of Combinatorial Optimization Problems Let the candidate sample pool be The budget for expanding the sample is The basic training set is The problem of optimizing the selection of augmented samples can be formalized as:

[0091] Where fitness function Defined as using augmented dataset Detection performance of the trained model on the validation set:

[0092] The feasible solution space of this problem is of size . .when , At this stage, the number of feasible solutions is excessively large, classifying it as a typical large-scale combinatorial optimization problem. Even with exhaustive search using high-performance computing equipment, the time required is still far beyond acceptable limits. The above scale analysis indicates that exact solution methods are computationally infeasible, necessitating the use of heuristic optimization strategies.

[0093] Proposition 2 (NP-hardness of the problem): The augmented sample selection problem is NP-hard.

[0094] Proof: This is proved by reduction. The classic Weighted Maximum Coverage Problem (WMCP) is reduced to this problem.

[0095] WMCP is defined as follows: Given the entire set Subset family (in ), element weight function and budget Please solve:

[0096] The reduction is constructed as follows: Complete Collection Discretized grid corresponding to the disturbance parameter space

[0097] Each candidate sample Corresponding subset This indicates the interference area that the sample can effectively improve. Element weight Corresponding grid points Sensitivity value

[0098] Sample selection budget Corresponding coverage budget

[0099] Under this reduction, choose Maximizing the model's robustness using a limited number of samples is equivalent to selecting... A subset is used to maximize the total weight of the covered elements. Since WMCP is NP-hard, the original problem is also NP-hard.

[0100] Corollary 2: Unless P=NP, there is no polynomial-time exact algorithm to solve the augmented sample selection problem. Therefore, using a heuristic optimization algorithm is a reasonable choice to solve this problem.

[0101] Algorithm selection criteria: This paper selects genetic algorithm as the solution strategy based on the following considerations: (1) Global search capability: Genetic algorithms achieve parallel exploration of the solution space by maintaining a population, which is less likely to get trapped in local optima compared to greedy algorithms; (2) Problem structure adaptability: The sample selection problem is naturally suitable for integer encoding representation, and the crossover and mutation operations of the genetic algorithm can be directly applied to the sample index sequence; 2. Genetic representation and population initialization strategies The performance of genetic algorithms is highly dependent on the encoding scheme and the design of the initial population. This invention addresses the characteristics of the augmented sample selection problem by proposing a customized genetic representation scheme and a strategy for generating a diverse initial population.

[0102] The genetic algorithm constructed in this paper adopts a fixed-length integer encoding scheme. One chromosome Indicated as containing The ordered sequence of one gene:

[0103] Each gene Indicates the index of the selected sample in the candidate pool. Chromosome. The corresponding augmented sample set is To avoid duplicate selection of the same sample, it is required that the genes within the chromosome are distinct, i.e. For all Established. Compared to binary encoding (using a length of...) The 0-1 vector representation selection scheme), integer encoding has the following advantages: (1) Indicating compactness: Chromosome length is fixed at 1000-1000 mm. Instead ,when Significantly reduces storage overhead; (2) Constraints are naturally satisfied: Budget constraints The encoding length implicitly guarantees that no additional repair mechanism is needed; (3) Efficiency of genetic operations: Crossover and mutation operations directly affect the sample index, with clear semantics and high efficiency.

[0104] The diversity of the initial population is crucial to the global search capability of the genetic algorithm. If the initial population is too similar, the algorithm is prone to getting trapped in local optima. In this problem, the candidate sample pool... The initial population is generated by aggregating key sampling points from multiple performance surfaces, and the number of samples for different types of interference may vary significantly. If uniform random sampling is used to generate the initial population, high-frequency interference types will be oversampled, resulting in uneven interference coverage of the initial solution. Let the sample The type of composite interference is Interference type The frequency in the candidate pool is defined as:

[0105] To promote balanced coverage of interference types, this paper proposes a sampling strategy based on inverse frequency weighting. (Samples) The selection weight is defined as:

[0106] in This is a smoothing factor to avoid numerical problems in the zero-frequency case. The normalized selection probability is:

[0107] in This is the current set of available samples. Proposition 3 (Guarantee of Balance in Inverse Frequency Weighting): Let the candidate sample pool be... Include Types of interference, type The number of samples is Total number of samples Under the inverse frequency weighting strategy, the probability of each interference type being sampled. satisfy:

[0108] in The minimum number of samples for each type. When At that time, the sampling probabilities of each type tend to be uniform. Proof: Let type The frequency is Then each type The selection weights for the samples are:

[0109] type The total weight is . The total global weight is:

[0110] type The probability of being sampled is:

[0111] (I) Limit Case Analysis: when hour, For all types is established, therefore:

[0112] That is, the sampling probabilities of each type tend to be uniformly distributed.

[0113] (II) Limited The following error bound: definition ,but . Note about Monotonically increasing, let . For any type The deviation of its sampling probability from the uniform distribution is:

[0114] because The maximum deviation occurs at extreme types:

[0115] calculate :

[0116] Substituting, we get:

[0117] 3. Fitness Assessment and Proxy Model The fitness function is a core component of genetic algorithms, and its design directly affects the correctness of the optimization direction. This invention discusses the definition of the fitness function and its efficient evaluation strategy.

[0118] chromosome The fitness value is defined as the detection performance of the augmented model on the validation set:

[0119] in For training on mixed training sets The model after the wheel, The average precision is the mean value for IoU thresholds ranging from 0.50 to 0.95. Fitness evaluation involves the complete model training and inference process, which has a significant computational cost. Let the time for a single training session be... Population size is The maximum number of generations is The total evaluation time is:

[0120] To reduce computational overhead, this paper adopts the following strategy: (1) Proxy Model Strategy: Use a lightweight detection model to replace the target model for fitness evaluation. The selection of the proxy model should meet the following criteria: efficiency, i.e., fast model training and inference speed, supporting rapid iterative evaluation; order preservation, i.e., the performance ranking of the proxy model is highly correlated with that of the target model. (2) Training set reduction strategy: A simplified basic training set is constructed using the proposed active learning strategy. This further reduces the cost per training session.

[0121] 4. Genetic manipulation design Genetic algorithms drive population evolution through three basic operations: selection, crossover, and mutation. This invention designs customized genetic operations specifically for the augmented sample selection problem. This paper employs a tournament selection strategy. Each time, samples are randomly drawn from the population. Individuals form a competitive set, and the one with the highest fitness is selected as the parent:

[0122] Tournament Scale Controlling selection pressure: The larger the value, the higher the probability of outstanding individuals being selected, the faster the convergence speed, but the lower the diversity. The smaller the value, the more random the selection becomes, which is beneficial for maintaining population diversity but slows down the convergence speed. Suppose that individuals in the population are arranged in descending order of fitness, and the rank is... The probability that an individual is selected is:

[0123] This probability varies with ranking. The rapid decline following the increase reflects a preference for outstanding individuals. A single-point crossover strategy is used. Given two parent chromosomes... and In the interval Randomly select intersection points Gene fragments exchanged at crossover points generate offspring:

[0124] Crossover operations with probability Execution. The core function of crossover is to combine the superior gene patterns of different parents to achieve efficient exploration of the solution space. A point mutation strategy is employed. Mutation is independently assessed for each gene locus on the chromosome, using probability... Replace it with a random index from the candidate pool:

[0125] in These are independent random numbers. The mutation operation introduces new genetic material to prevent premature population convergence. Crossover and mutation operations may lead to duplicate genes in chromosomes, violating the distinctness constraint. This paper designs a deterministic repair mechanism: Let chromosomes There are duplicate gene sets in it. For each repeating position ,Will Replace with random samples that were not selected from the candidate pool:

[0126] This repair mechanism ensures that chromosomes always meet the constraints after genetic manipulation.

[0127] 5. Population Renewal and Convergence Analysis To prevent genetic operations from corrupting the current optimal solution, an elite preservation strategy is employed. In each generation of evolution, the top-fittest individuals are selected... Individuals are directly replicated into the next generation of the population, among which This refers to the elite retention rate. Under an elite retention strategy, the optimal fitness value of the population monotonically does not decrease with each generation, i.e.:

[0128] in . Proof: Let the first... The optimal individual in the generation is The strategy of retaining elites It was directly copied to the first page. In the generational population. Therefore:

[0129] Elite retention rate With the rate of variation Together, they determine the diversity-convergence tradeoff. Let... For the first Regarding population diversity, then:

[0130] when When it is too big, When the population tends to be homogeneous, the algorithm is prone to getting trapped in local optima; when When the size is too large, the optimal gene pattern is frequently disrupted, and the convergence efficiency decreases. Termination condition: The algorithm terminates when any of the following conditions are met: (1) Upper limit of generation: The number of generations of evolution reaches the preset maximum value. ; (2) Convergence criterion: Continuous The improvement in optimal fitness within a generation is less than the threshold. :

[0131] The convergence criterion is based on the following observation: when the population is close to the optimal solution, the rate of fitness improvement tends to level off, and the marginal benefit of further evolution diminishes.

[0132] After the algorithm terminates, it outputs the optimal chromosome that appears during the evolutionary process and its corresponding augmented sample set:

[0133] Basic training set reduction for computational efficiency The genetic algorithm described in the previous section requires fitness evaluation of a large number of chromosomes during the iterative process, and each evaluation involves a complete model training process. When the original training set is large, computational overhead becomes a key bottleneck restricting the practicality of the algorithm. From the perspective of information theory and active learning, this invention proposes a basic training set reduction strategy, aiming to construct a representative sample subset that is smaller in size but retains information, thereby improving the computational efficiency of the genetic algorithm while ensuring model performance. The framework is shown in Figure 4.

[0134] 1. Computational bottleneck analysis and problem formalization Let the population size of the genetic algorithm be... The maximum number of generations is The original training set size is The training time for a single sample is The number of training rounds is The total computation time of the genetic algorithm is:

[0135] This analysis shows that, It is one of the main contributors to computational overhead. If the size of the training set can be increased from... Reduced to The calculation time will be reduced accordingly to the original time. times. Let the original training set be ,in For image, The corresponding object detection labels are provided. The training set reduction problem can be formalized as:

[0136] in For the target size, For the information loss function, the metric uses subsets. The trained model is relative to using the full set Performance gap between trained models:

[0137] in Indicates in the dataset The model trained on it, To predict the measure of difference. Directly optimizing the objective function is computationally infeasible because it evaluates arbitrary subsets. The information loss requires a complete model training process. Therefore, it is necessary to design a surrogate criterion that can be efficiently computed to approximate the information value of the samples. 2. Measurement of the value of sample information based on uncertainty Active learning theory suggests that samples with high model prediction uncertainty are typically located near the decision boundary and have higher informational value for model training. This invention establishes a theoretical link between uncertainty and sample value from an information theory perspective and designs a specific uncertainty metric for object detection tasks.

[0138] Model For the sample The predicted distribution is ,sample The value of information can be measured by expected information gain:

[0139] in It is the entropy function. Intuitively, EIG measures the entropy of observed samples. The expected reduction in model uncertainty after obtaining the true labels. High EIG samples contribute more to eliminating model uncertainty and therefore have higher training value. However, directly calculating EIG in object detection tasks faces two difficulties: (1) the output space contains the joint distribution of bounding boxes and categories, making entropy calculation complex; and (2) it requires integration over all possible labels, resulting in high computational overhead. To address these challenges, this paper proposes an approximate uncertainty measure based on detection confidence.

[0140] Model For the sample The test results are ,in For bounding box, For classification confidence, To determine the number of targets to be detected, define the sample. The overall uncertainty score is:

[0141] in These are the weighting coefficients. and These are the confidence level variance term and the low confidence level proportion term, respectively. Confidence variance term: reflects the internal consistency of the test results.

[0142]

[0143] in The standard deviation is represented by . High variance indicates that the model makes significant discrepancies in its judgments of different objects in the same image, suggesting that the sample contains complex features that the model has difficulty processing consistently. Low confidence percentage: Reflects the proportion of difficult targets.

[0144]

[0145] in The low confidence threshold This is an indicator function. This measure measures the model's overall grasp of the targets in the sample; a high proportion indicates that the sample contains more challenging targets. If the model is in the sample No target was detected. This indicates that the sample poses an extreme challenge to the current model and should be assigned the highest level of uncertainty.

[0146] This means the model's recall is zero for this sample, indicating a serious false negative problem. Including such samples in the training set can help improve the model's ability to detect difficult scenarios. Uncertainty measures based on confidence level under the mild assumption. With expected information gain Positive correlation. Proof: Classification confidence This can be viewed as a point estimate of the class posterior probability by the model. When The entropy of the category distribution reaches its maximum when it approaches 0.5; when When the entropy approaches 0 or 1, it approaches zero. Therefore, low-confidence detection corresponds to high-entropy prediction, and the confidence variance reflects the degree of entropy variation within the sample. The weighted combination of the two approximately characterizes the overall information entropy of the sample and is positively correlated with EIG.

[0147] 3. Joint constraints of category balance and difficulty perception Simply selecting samples based on uncertainty may lead to two problems: (1) class distribution imbalance, with rare class samples being ignored; (2) difficulty distribution skew, with too many difficult samples affecting convergence stability. This invention proposes a joint constraint mechanism that integrates class balance and difficulty awareness to ensure that the reduced dataset maintains good statistical properties.

[0148] Suppose the current basic training set Medium category The number of annotations is The total number of annotations is Then the category The annotation ratio is:

[0149] In each round of sample selection, a selection quota proportional to the current proportion is allocated to each category. Let the target selection quantity for this round be... Then the category The quota is:

[0150] in This is the floor function. Ensure each category receives at least one quota. Reduce the dataset under the category balance constraint. The category distribution of the original dataset The category distributions are asymptotically consistent.

[0151] Let the original dataset contain categories The proportion is During the iterative selection process, the category... The cumulative number of selections is:

[0152] If the class distribution of the initial seed set is close ,but For all Established, thus .

[0153] Based on the detection performance of the pre-trained model, the samples are divided into three difficulty levels: Easy: All targets were detected with high confidence. ) Medium: A medium confidence level of detection exists. ) Hard: Low-confidence detections or missed detections exist. or ) in , The confidence threshold. This represents the actual target quantity. During the selection process within each category, stratified sampling is performed according to a preset difficulty distribution:

[0154] 4. Iterative Refinement Strategy The training set reduction is achieved through multiple iterative rounds, with each round comprising two stages: model update and sample selection. The core idea is that as the model's capabilities improve, the uncertainty distribution changes, necessitating dynamic adjustment of sample selection to adapt to the model's current state. Each iteration selects the most uncertain sample for the current model and adds it to the training set. This gradually reduces the model's uncertainty region, and the accumulated information gain monotonically increases and is bounded with the number of iterations.

[0155] The iterative process terminates when any of the following conditions are met: (1) The size of the reduced set reaches the target value. (2) If the average uncertainty of newly added samples in several consecutive rounds is lower than the preset threshold, it indicates that the information tends to be saturated and the marginal benefit of continuing to expand the training set is diminishing.

[0156] Iterative refinement requires an initial seed set. As a starting point, its size needs to meet two constraints: firstly, it should be sufficient to train an initial model with basic detection capabilities, making uncertainty assessment meaningful; secondly, it should be much smaller than the target size, leaving ample room for subsequent iterations. This paper uses hierarchical random sampling to ensure that the seed set covers all target categories and has a balanced difficulty distribution.

[0157] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which is executed by a processor to implement a robust augmented sample selection method for target detection based on a two-dimensional performance surface.

[0158] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement a robust augmented sample selection method for target detection based on a two-dimensional performance surface.

[0159] The above description is merely a preferred embodiment of a robust augmented sample selection method for target detection based on two-dimensional performance surfaces. The scope of protection for such a method is not limited to the above embodiments; all technical solutions falling within this conceptual framework are within the scope of protection of this invention. It should be noted that for those skilled in the art, any improvements and variations made without departing from the principles of this invention should also be considered within the scope of protection of this invention.

Claims

1. A robust augmented sample selection method for target detection based on two-dimensional performance surfaces, characterized by: The method includes the following steps: Step 1: Construct a robust augmentation sample candidate pool and select the candidate samples that are most valuable for improving the robustness of the model; Step 2: Optimize the selection of augmented samples based on genetic algorithm, and output the augmented sample set corresponding to the chromosome with the highest fitness; Step 3: Based on active learning, reduce the basic training set by selecting a representative subset of samples from the original training set to construct the basic training set; Step 4: Perform robustness enhancement training on the model, construct a robustness enhancement training set, and train the object detection model.

2. The method according to claim 1, characterized in that: set up For object detection models, For the test set, and There are two types of interference, and their intensity parameters are as follows: and The two-dimensional performance response function is defined as follows: in For two-dimensional interference parameter space, To the interference parameters The generated interference test set; Performance response function A two-dimensional surface is formed in the parameter space, and its geometric shape reflects the model's response characteristics to different combinations of disturbances; Four representative types of interference in remote sensing imaging were selected, namely: , , , ; Based on these four types of interference, four sets of composite interference scenarios are constructed by combining them in pairs: Let the performance response function be defined. At point If the model is continuously differentiable at a point, its local robustness at that point is defined as follows: in, It is a unit direction vector. This is a preset performance tolerance threshold; The measurement of performance degradation exceeding [a certain value] in the most unfavorable direction was taken. Minimum required disturbance amount; Let the performance surface There are a total of Given a set of grid points, sort all grid points in descending order of sensitivity. Let the sorted sensitivity sequence be denoted as . ,in The cumulative sensitivity contribution rate function is defined as: The function describes the order of sorting before and after sorting. The contribution percentage of each highly sensitive point to the total sensitivity, given a target contribution rate. Number of key sampling points Defined as the cumulative contribution rate reaching a certain level for the first time. Minimum number of points: The corresponding adaptive sensitivity threshold is ,parameter It controls the trade-off between the size and coverage of the candidate sample pool; For performance surfaces The set of key sampling points is defined as follows: Each key sampling point Corresponding to a set of samples under specific combined interference conditions Its size is equal to the number of images in the original test set. ; For a single performance surface, its candidate sample set is: The candidate sample pool is formed by integrating four sets of composite interference scenarios: Assume that the average selection for each performance surface is... If there are 1 key sampling point, then the size of the candidate sample pool satisfies: The equality holds when there is no overlap at the key sampling points of each surface.

3. The method according to claim 2, characterized in that: set up Candidate sample pool is The budget for expanding the sample is The basic training set is The problem of augmenting the sample selection is formalized as follows: Wherein, fitness function Defined as using augmented dataset Detection performance of the trained model on the validation set: The size of the feasible solution space of the problem is ;when , In such cases, heuristic optimization strategies should be employed.

4. The method according to claim 3, characterized in that: The constructed genetic algorithm uses fixed-length integer encoding, with one chromosome... Indicated as containing The ordered sequence of one gene: Each gene Indicates the index of the selected sample in the candidate pool; chromosome The corresponding augmented sample set is To avoid repeated selection of the same sample, it is required that the genes within the chromosome are distinct, i.e. For all Established; Let the sample The type of composite interference is Interference type The frequency in the candidate pool is defined as: sample The selection weight is defined as: in To act as a smoothing factor and avoid numerical problems in the zero-frequency case; the normalized selection probability is: in This is the current set of available samples; chromosome The fitness value is defined as the detection performance of the augmented model on the validation set: in, For training on mixed training sets The model after the wheel, The mean precision is the average value of the IoU threshold ranging from 0.50 to 0.

95. Let the training time be... Population size is The maximum number of generations is The total evaluation time is: A tournament selection strategy is used, with random selection from the population each time. Individuals form a competitive set, and the one with the highest fitness is selected as the parent: Suppose that individuals in the population are arranged in descending order of fitness, and the rank is... The probability of an individual being selected is: Using a single-point crossover strategy, given two parent chromosomes... and In the interval Randomly select intersection points Gene fragments exchanged at crossover points generate offspring: Crossover operations with probability The core function of execution and crossover operations is to combine the superior gene patterns of different parents to achieve efficient exploration of the solution space; A point mutation strategy is employed, where each gene locus in the chromosome is independently assessed for mutation, using probability. Replace it with a random index from the candidate pool: in, These are independent random numbers.

5. The method according to claim 3, characterized in that: Design a deterministic repair mechanism: Let chromosomes There are duplicate gene sets in it. For each repeating position ,Will Replace with random samples that were not selected from the candidate pool: Repair mechanisms ensure that chromosomes always meet the constraints after genetic manipulation.

6. The method according to claim 5, characterized in that: In each generation of evolution, the fitness ranking is determined by the number of generations. Individuals are directly replicated into the next generation of the population. Let be the elite retention rate. Under an elite retention strategy, the optimal fitness value of the population monotonically does not decrease with each generation, i.e.: in ; Let the first The optimal individual in the generation is The strategy of retaining elites It was directly copied to the first page. In the population; therefore: Elite retention rate With the rate of variation Together, they determine the diversity-convergence tradeoff; let... For the first Regarding population diversity, then: when When it is too big, When the population tends to be homogeneous, the algorithm is prone to getting trapped in local optima; when When the size is too large, the optimal gene pattern is frequently disrupted, and the convergence efficiency decreases. Termination condition: The algorithm terminates when any of the following conditions are met: (1) Upper limit of generation: The number of generations of evolution reaches the preset maximum value. ; (2) Convergence criterion: Continuous The improvement in optimal fitness within a generation is less than the threshold. : The convergence criterion is based on the following observation: when the population is close to the optimal solution, the rate of fitness improvement tends to level off, and the marginal benefit of further evolution diminishes; After the algorithm terminates, it outputs the optimal chromosome that appears during the evolutionary process and its corresponding augmented sample set: 。 7. The method according to claim 6, characterized in that: set up The population size of the genetic algorithm is The maximum number of generations is The original training set size is The training time for a single sample is The number of training rounds is The total computation time of the genetic algorithm is: Let the original training set be ,in For image, For the corresponding object detection annotations, the training set reduction problem is formalized as follows: in, For the target size, For the information loss function, the metric uses subsets. The trained model is relative to using the full set Performance gap between trained models: in, Indicates in the dataset The model trained on it, To predict the measure of difference; Set up a model For the sample The predicted distribution is ,sample The value of information can be measured by expected information gain: in, EIG is the entropy function; it measures the observed samples. The expected reduction in model uncertainty after the true labels are obtained; Set up a model For the sample The test results are ,in For bounding box, For classification confidence, To determine the number of targets to be detected, define the sample. The overall uncertainty score is: in, These are the weighting coefficients. and These are the confidence level variance term and the low confidence level proportion term, respectively. When the model is in the sample No target was detected. This indicates that the sample poses an extreme challenge to the current model and should be assigned the highest level of uncertainty: Uncertainty measures based on confidence level under the mild assumption. With expected information gain Positive correlation; Suppose the current basic training set Medium category The number of annotations is The total number of annotations is Then the category The annotation ratio is: In each round of sample selection, a selection quota proportional to the current proportion is allocated to each category; let the target selection quantity for this round be... Then the category The quota is: in, This is the floor function. Ensure each category receives at least one quota; reduce the dataset under the category balance constraint. The category distribution and the original dataset The category distributions are asymptotically consistent; Let the categories in the original dataset be... The proportion is During the iterative selection process, the category The cumulative number of selections is: If the class distribution of the initial seed set is close ,but For all Established, thus .

8. The method according to claim 7, characterized in that: Based on the detection performance of the pre-trained model, the samples are divided into three difficulty levels: Simple: All targets were detected with high confidence. ; Medium: Medium confidence level detection exists. ; Challenges: Low-confidence detection or missed detections exist. or ; in, , The confidence threshold. This represents the actual target quantity. During the selection process within each category, stratified sampling is performed according to a preset difficulty distribution: The iterative process terminates when either of the following conditions is met: the size of the reduced set reaches the target value. If the average uncertainty of newly added samples is lower than the preset threshold within several consecutive rounds, it indicates that the information is approaching saturation, and the marginal benefit of continuing to expand the training set is diminishing.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, The program is executed by the processor to implement the method as claimed in any one of claims 1-8.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the method of any one of claims 1-8.

Images