Incremental learning-supported computer vision object recognition methods and systems

Abstract

This invention relates to a computer vision target recognition method and system supported by incremental learning, belonging to the field of image processing. It obtains associated images and a first high-frequency background image, as well as an image to be recognized and a second high-frequency background image, from a target detection model. The method calculates a first semantic similarity between the associated image and the image to be recognized, a second semantic similarity between the first and second high-frequency background images, and a distribution similarity between a first distribution feature of the associated image in the first high-frequency background image and a second distribution feature of the image to be recognized in the second high-frequency background image. Based on the first and second semantic similarities and the distribution similarity, a task similarity analysis is performed to obtain a task similarity. When the similarity is greater than or equal to a similarity threshold, incremental learning of the model using a shared backbone network is performed; otherwise, incremental learning of branch networks is performed. This method solves the problem of insufficient flexibility and accuracy in target recognition under dynamic environments, leading to false detections or missed detections.

Description

Technical Field

[0001] This invention relates to the field of image processing, and more particularly to a computer vision target recognition method and system supported by incremental learning. Background Technology

[0002] In the field of computer vision, object recognition is a crucial task, requiring systems to accurately identify and classify target objects in images. Traditional object recognition methods typically rely on training with large amounts of labeled data to build accurate object detection models. However, in practical applications, labeled data is often difficult to obtain, and the target scene and background may change over time. This necessitates that object detection models can adapt to these changes, i.e., possess the ability to learn incrementally.

[0003] Incremental learning is a method of gradually updating a model to adapt to new data, integrating new knowledge into an existing model without retraining the entire model. However, traditional incremental learning methods in computer vision object recognition primarily focus on the features of the object itself, such as category and size, while neglecting the importance of background features. In dynamic environments, changes in the background can significantly impact object detection. For example, when switching between urban street scenes and outdoor scenes, traditional incremental learning methods may lead to false positives or false negatives due to significant differences in background features such as buildings, vegetation, and sky. Furthermore, traditional incremental learning for object detection typically employs a fixed-modality learning strategy, rather than flexibly selecting strategies based on the specific characteristics of the task, resulting in inflexible and inaccurate model update strategies. Summary of the Invention

[0004] This invention addresses the technical problem of insufficient flexibility and accuracy in target recognition under dynamic environments in existing technologies, which leads to false detections or missed detections. It provides a computer vision target recognition method and system supported by incremental learning to solve this problem.

[0005] The technical solution of the present invention to solve the above-mentioned technical problems is as follows:

[0006] In a first aspect, the present invention provides a computer vision target recognition method supported by incremental learning, comprising: obtaining an associated target image and a first high-frequency background image of a basic target detection model, as well as a target image to be recognized and a second high-frequency background image; calculating a first semantic similarity between the associated target image and the target image to be recognized; calculating a second semantic similarity between the first high-frequency background image and the second high-frequency background image; calculating the distribution similarity between a first distribution feature of the associated target image in the first high-frequency background image and a second distribution feature of the target image to be recognized in the second high-frequency background image; performing task similarity analysis based on the first semantic similarity, the second semantic similarity, and the distribution similarity to obtain a task similarity; when the task similarity is greater than or equal to a task similarity threshold, performing incremental learning of a shared backbone network on the basic target detection model for target recognition; otherwise, performing incremental learning of an additional branch network on the basic target detection model for target recognition.

[0007] Secondly, the present invention provides a computer vision target recognition system supported by incremental learning, comprising: an image recognition module for obtaining associated target images and a first high-frequency background image of a basic target detection model, as well as a target image to be recognized and a second high-frequency background image; a first similarity calculation module for calculating a first semantic similarity between the associated target image and the target image to be recognized; a second similarity calculation module for calculating a second semantic similarity between the first high-frequency background image and the second high-frequency background image; a distribution similarity calculation module for calculating the distribution similarity between a first distribution feature of the associated target image in the first high-frequency background image and a second distribution feature of the target image to be recognized in the second high-frequency background image; a task analysis module for performing task similarity analysis based on the first semantic similarity, the second semantic similarity, and the distribution similarity to obtain a task similarity; and a target recognition module for performing target recognition by incremental learning of a shared backbone network on the basic target detection model when the task similarity is greater than or equal to a task similarity threshold, otherwise performing target recognition by incremental learning of an additional branch network on the basic target detection model.

[0008] The beneficial effects of this invention are: by comprehensively considering the semantic similarity and distribution characteristics of the target image and the background image, task similarity analysis is performed, and incremental learning strategies such as sharing the backbone network or adding branch networks are flexibly selected based on the analysis results, thereby improving the accuracy and adaptability of computer vision target recognition. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating the incremental learning-supported computer vision target recognition method provided by the present invention.

[0010] Figure 2 A schematic diagram of the structure of the computer vision target recognition system supported by incremental learning provided by the present invention.

[0011] Explanation of reference numerals in the attached figures: Image recognition module 11, First similarity calculation module 12, Second similarity calculation module 13, Distribution similarity calculation module 14, Task analysis module 15, Target recognition module 16. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0013] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0014] In the description of this invention, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this invention is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed herein.

[0015] Example 1:

[0016] like Figure 1 As shown, this embodiment of the invention provides a computer vision target recognition method supported by incremental learning, including:

[0017] S10: Obtain the associated target image and the first high-frequency background image of the basic target detection model, as well as the target image to be identified and the second high-frequency background image.

[0018] S20: Calculate the first semantic similarity between the associated target image and the target image to be identified.

[0019] S30: Calculate the second semantic similarity between the first high-frequency background image and the second high-frequency background image.

[0020] S40: Calculate the distribution similarity between the first distribution feature of the associated target image in the first high-frequency background image and the second distribution feature of the target image to be identified in the second high-frequency background image.

[0021] S50: Perform task similarity analysis based on the first semantic similarity, the second semantic similarity, and the distribution similarity to obtain task similarity.

[0022] S60: When the task similarity is greater than or equal to the task similarity threshold, the basic target detection model is subjected to incremental learning of the shared backbone network for target recognition; otherwise, the basic target detection model is subjected to incremental learning of the branch network for target recognition.

[0023] For example, the basic object detection model consists of a backbone network, a feature fusion network (neck), and a detection head. The backbone network is responsible for extracting image features, the feature fusion network is used to further integrate and enhance these features, and the detection head outputs the location and category information of the target. In the context of incremental learning, the specific model structure is dynamically adjusted according to task requirements. For incremental learning with a shared backbone network, the new task shares the backbone network part of the original model, only fine-tuning the detection head or specific layers. For incremental learning with added branch networks, the model adds new branch networks after the backbone network to adapt to the needs of the new task. The model input includes the associated target image and the first high-frequency background image, the target image to be identified, and the second high-frequency background image. The associated target image and the first high-frequency background image come from the training set or validation set of the basic object detection model and are used to compare with the target image to be identified to evaluate task similarity. The associated target image is an image containing a known target object, while the first high-frequency background image is the background part of these images, which has been processed by high-frequency filtering to highlight background features. The target image to be identified and the second high-frequency background image are the targets that the model needs to identify, both containing the target object and the background. The second high-frequency background image is the background portion of the target image to be identified, processed by high-frequency filtering. The model's output includes task similarity and recognition result. Task similarity is calculated by comparing the semantic and distributional features of the associated target image and the target image to be identified. This value is used to determine the similarity between the current task and the tasks already learned by the basic object detection model. The recognition result refers to the target image to be identified, after the model selects an appropriate incremental learning strategy based on the task similarity, and the model outputs the recognition result, including the location and category information of the target object. By comprehensively considering the semantic similarity and distributional features of the target image and the background image, the model can more accurately evaluate task similarity and thus select a suitable incremental learning strategy. This helps the model quickly adapt to new tasks while maintaining existing knowledge, improving recognition accuracy. Simultaneously, incremental learning allows the model to continuously learn new knowledge while retaining previously learned information. This dynamic adaptability enables the model to cope with constantly changing data environments, improving its applicability in different scenarios. By flexibly choosing incremental learning strategies that involve sharing a backbone network or adding branch networks, the consumption of computing resources and storage space can be reduced without sacrificing performance. This helps to lower the deployment cost of the model and improve its practicality. In summary, by comprehensively considering multiple factors in task similarity analysis and selecting appropriate incremental learning strategies based on the analysis results, the model improves the accuracy and adaptability of target recognition while optimizing resource utilization.

[0024] After acquiring the associated target image and the first high-frequency background image, as well as the target image to be identified and the second high-frequency background image, the focus is on calculating similarity in two main aspects: first, the semantic similarity of the target object itself in the two images (first semantic similarity); and second, the high-frequency feature similarity of the background in the two images (second semantic similarity). These two similarity calculations allow for a more comprehensive assessment of the similarity between the two images in terms of content and background. First, two images are selected for comparative analysis: one is the known associated target image, and the other is the target image to be identified. To quantify the similarity between these two images in terms of content, an advanced semantic similarity calculation algorithm is used. This algorithm can deeply understand the visual features in the image, including but not limited to the shape, color, texture of objects, and their spatial relationships, thus providing a numerical value reflecting the content similarity between the two, i.e., the first semantic similarity. This step can preliminarily determine whether the image to be identified matches the known image. Next, to more accurately capture high-frequency details in the background, such as texture and edge information, high-frequency background extraction is performed on both the associated target image and the target image to be identified. High-frequency background images highlight rapidly changing parts of the image and are often closely related to the image's detailed features. Then, the semantic similarity calculation algorithm is used again to compare the extracted first and second high-frequency background images to obtain the second semantic similarity. This step helps to understand the consistency or difference between the two images in terms of background details more deeply, which is particularly important for certain application scenarios, such as background replacement and image synthesis. In summary, by calculating the first semantic similarity between the associated target image and the target image to be identified, as well as the second semantic similarity between their respective high-frequency background images, a comprehensive and in-depth analysis of the similarity between the two images in terms of content and background details can be achieved, providing strong support for subsequent image recognition, classification, or synthesis tasks.

[0025] Subsequently, feature extraction algorithms are used to calculate the first distribution features presented by the combination of the associated target image and the first high-frequency background image, and the second distribution features presented by the combination of the target image to be identified and the second high-frequency background image. These distribution features not only reflect the attributes of the target object itself, but also incorporate detailed information from the background environment, thus forming a more comprehensive image feature description. An efficient similarity calculation algorithm is used to compare the first and second distribution features to obtain the distributional similarity between them. The key to this step is the ability to accurately measure the degree of similarity in the distribution of the two sets of features, thus providing a strong basis for judging the correlation between the target image and the image to be identified. This method can more accurately assess the overall consistency between the target image and the background image, thereby making more accurate judgments in target recognition tasks. It not only improves the accuracy of target recognition but also enhances the model's adaptability to complex background environments.

[0026] Furthermore, a comprehensive task similarity analysis is performed based on these three similarity metrics. This process not only considers the similarity of the target objects themselves but also deeply analyzes the impact of the background environment on task similarity, thus deriving a comprehensive task similarity value. This value is a quantitative indicator used to evaluate the similarity between the current task to be identified and the tasks already learned by the basic object detection model. To adaptively select a model update strategy, a task similarity threshold is set as the basis for judging the model update method. If the calculated task similarity is greater than or equal to this threshold, it indicates that the current task and the learned task have high similarity in both content and background. Therefore, incremental learning with a shared backbone network can be selected for the basic object detection model. This approach can fully utilize the model's existing feature extraction capabilities, only fine-tuning the detection head or specific layers to quickly adapt to new tasks while maintaining the recognition accuracy for old tasks. Conversely, if the calculated task similarity is lower than the threshold, it indicates that the current task and the learned task have significant differences in content or background. In this case, incremental learning with a shared backbone network may lead to a decrease in model performance. Therefore, incremental learning with added branch networks is selected for the basic object detection model. This approach allows the model to retain its original feature extraction capabilities while learning specific features of new tasks through newly added branch networks, thereby achieving accurate recognition of new tasks. In summary, by comprehensively considering the semantic and distributional similarity between the target and background images for task similarity analysis, and intelligently selecting the model update method based on the analysis results, the object detection model can efficiently learn new tasks while maintaining its ability to recognize old tasks during incremental learning, thus improving recognition accuracy and adaptability.

[0027] In a preferred embodiment, obtaining the associated target image and the first high-frequency background image of the basic target detection model includes:

[0028] Obtain the first historical background image up to the Nth historical background image of the associated target image; perform semantic similarity clustering on the first historical background image up to the Nth historical background image to obtain multiple groups of historical background images; randomly select one image from each class whose number of images in the group is greater than or equal to the threshold of the number of images in the group and add it to the first high-frequency background image.

[0029] Optionally, associated target images and a series of historical background images related to the basic target detection model are collected. These historical background images represent the background environment of the target at different times or under different environments, providing rich background information for the model. Specifically, the first historical background image of the associated target image is retrieved from the database, up to the Nth historical background image, where N is a preset integer representing the upper limit of the number of historical background images considered. Subsequently, in order to extract representative high-frequency background images from these historical background images, a semantic similarity clustering method is adopted. This method is based on the semantic content of the images and groups images with high similarity into one category. By calculating the semantic similarity between each historical background image and other images, a similarity matrix can be constructed, and clustering algorithms (such as K-means, DBSCAN, etc.) are applied to identify different image groups. After clustering, multiple sets of historical background images are obtained. Next, the most representative images need to be selected from these groups to form a high-frequency background image set. The selection criterion is the number of images within a group: if the number of images in a certain class is greater than or equal to a preset threshold for the number of images within a group, it means that the images in that class have high similarity and representativeness. For such a class, an image is randomly selected and added to the first high-frequency background image set. This process ensures that even with a large number of historical background images, the most representative image can be efficiently extracted for subsequent analysis and model training. Similarly, when it is necessary to determine the second high-frequency background image for the target image to be identified, the same determination logic as for the first high-frequency background image is followed. That is, a series of historical background images of the target image to be identified are collected, semantic similarity clustering is performed, and the most representative image is selected according to the same criteria to form the second high-frequency background image set. In summary, this process, through careful selection and processing of associated target images and their historical background images, helps to improve the generalization ability and recognition accuracy of the target detection model. At the same time, by adopting semantic similarity clustering and representative image selection methods, it ensures that even with a large amount of image data, the most valuable information can be extracted efficiently.

[0030] In a preferred embodiment, semantic similarity clustering is performed on the first historical background image up to the Nth historical background image to obtain multiple sets of historical background images, including:

[0031] Perform pairwise semantic similarity calculations on the first historical background image up to the Nth historical background image to obtain several semantic similarities; extract k semantic similarities from the several semantic similarities that are closest to the first historical background image, set as the first adjacent semantic similarity, and calculate the mean to obtain the first distribution density; continue until the Nth distribution density is obtained; calculate the mean density of the first distribution density up to the Nth distribution density, and compare it with the first distribution density up to the Nth distribution density to obtain the first outlier factor up to the Nth outlier factor; based on the outlier factor threshold, and combined with the first outlier factor up to the Nth outlier factor, cluster the first historical background image up to the Nth historical background image to obtain the multiple groups of historical background images.

[0032] Specifically, a series of historical background images related to the base object detection model are collected, from the first historical background image to the Nth historical background image. These images constitute the original dataset for cluster analysis. Then, each pair of historical background images is compared pairwise to calculate their semantic similarity. This process generates several semantic similarity values, each representing the degree of similarity between the two images. With these semantic similarity values, further processing is performed to extract the "adjacent semantic similarity" for each image. Specifically, for the first historical background image, the k nearest to furthest semantic similarities are extracted—that is, the similarities to the k most similar images. The mean of these similarities is then calculated to obtain the first distribution density for the first image. The same process is repeated for all other historical background images until the Nth distribution density is calculated for the Nth historical background image. Now, a set of distribution densities for each image is obtained. To evaluate the outlierness of each image relative to the entire dataset, the mean density of this set is calculated, and the distribution density of each image is compared to this mean density to obtain an outlier factor. The outlier factor is a relative indicator that reflects the degree of isolation of each image in the semantic space. After obtaining the outlier factor, historical background images can be clustered according to a preset outlier factor threshold. Specifically, images with an outlier factor less than the threshold are considered "normal" images, while images with an outlier factor greater than or equal to the threshold are considered "outliers." Then, various clustering algorithms (such as K-means, hierarchical clustering, etc.) can be used, combined with the information from these outlier factors, to divide the historical background images into multiple groups. Images within each group are relatively similar semantically, while images between different groups show significant differences. In summary, by calculating pairwise semantic similarity, extracting adjacency semantic similarity, calculating distribution density and outlier factor of historical background images, and combining this with outlier factor thresholds for clustering, a series of historical background images are divided into multiple groups with semantic similarity. This process provides valuable background information for subsequent object detection model training, helping to improve the model's generalization ability and recognition accuracy.

[0033] In a preferred embodiment, the formula for calculating the first semantic similarity is:

[0034] in, The feature covariance matrix of the target image X at the k-th layer of the feature extraction network. The feature covariance matrix ω of the target image Y in the k-th layer of the feature extraction network. k The weight coefficients represent the features of the k-th layer, K represents the number of feature layers selected, Tr(·) represents the matrix trace operation, and represents the energy of the covariance matrix.

[0035] Furthermore, traditional similarity calculation methods, such as cosine similarity, mainly rely on the alignment of the mean of feature vectors, which can reflect the overall similarity between images to a certain extent. However, this method ignores the correlation between feature channels, i.e., the statistical distribution information such as texture and shape in the image, which is crucial for a deeper understanding of image content. Therefore, to overcome this limitation, this application designs a new semantic similarity calculation formula, which measures the similarity of target images in local details (shallow features) and semantic abstraction (deep features) through multi-level covariance alignment. The motivation for this design is to fully utilize the information of each layer of the feature extraction network to capture the correlation between image feature channels, thereby more comprehensively evaluating the semantic similarity between images. Specifically, the first semantic similarity calculation formula is as follows: in, The feature covariance matrix of the target image X at the k-th layer of the feature extraction network. The feature covariance matrix ω of the target image Y in the k-th layer of the feature extraction network. k The weight coefficients representing the features of the k-th layer are used to balance the importance of features from different layers. K represents the number of feature layers selected, i.e., the number of layers in the feature extraction network. Tr(·) represents the matrix trace operation, used to calculate the sum of the diagonal elements of the matrix, here used to measure the energy or overall variability of the covariance matrix. The physical meaning of this formula is that by calculating the feature covariance matrices of the target image at different layers of the feature extraction network and then weighting and summing them, a similarity measure that integrates local image details and semantic abstraction information can be obtained. This process not only considers the mean of the feature vectors but also delves into the correlation between feature channels, thus more accurately reflecting the semantic similarity between images. The advantage of this formula is its stronger robustness to target deformation and viewpoint changes. Because it combines multi-layer information from the feature extraction network, it can still capture stable semantic features even if the target changes in shape or viewpoint. This is suitable for fine-grained incremental learning scenarios, i.e., continuously learning and recognizing new target categories in a constantly updated dataset while maintaining accurate recognition of existing categories. In summary, the first semantic similarity calculation formula achieves a similarity measurement of target images in terms of local details and semantic abstraction through multi-level covariance alignment. This method not only overcomes the limitations of traditional similarity calculation methods but also demonstrates robustness to complex situations such as target deformation and viewpoint changes, providing a new solution for tasks such as fine-grained incremental learning.

[0036] In a preferred embodiment, the formula for calculating the second semantic similarity is:

[0037] Among them, I bg1The first high-frequency background image is characterized by extraction using a Butterworth high-pass filter. bg2 The second high-frequency background image is represented by Butterworth high-pass filtering. H(·) represents the multi-directional gradient histogram of the image block. The image is divided into B blocks, and an 8-directional gradient histogram is calculated for each block. ||·||1 represents the L1 norm and measures the histogram difference.

[0038] In detail, background images typically contain rich edge and texture structure information, which is crucial for understanding image content. However, directly using pixel differences to measure the similarity between background images is often highly sensitive to noise, especially in high-frequency components. To overcome this limitation, gradient histograms can be used to characterize the local structure distribution of an image. Gradient histograms not only capture the edge and texture information of an image but also exhibit robustness to changes in illumination and minor perturbations. Therefore, this application designs a second semantic similarity calculation formula, aiming to quantify the similarity between background images through the difference in gradient histograms. The second semantic similarity calculation formula is as follows: In the second semantic similarity calculation formula, two high-frequency background images extracted using Butterworth high-pass filtering were used, denoted as I... bg1 and I bg2The Butterworth high-pass filter effectively extracts high-frequency components from images, namely edge and texture information, while suppressing low-frequency noise. Next, the two high-frequency background images are divided into B blocks, and an 8-directional gradient histogram, denoted as H(·), is calculated for each block. This yields the gradient histogram representations of the two images in B blocks. To measure the difference between these two gradient histograms, the L1 norm is used as a metric. The L1 norm intuitively reflects the degree of difference between the two histograms in various directions. By calculating the L1 norm difference between these two histograms, a quantified measure of background structural similarity is obtained. The physical meaning of the second semantic similarity calculation formula lies in its exponential decay of the statistical differences in the multi-directional gradient histograms of image blocks to quantify the similarity of background structures. Specifically, by dividing the image into multiple blocks and calculating the gradient histogram of each block, edge and texture information in various local regions of the image can be captured. Then, by calculating the differences between these histograms, the structural similarity of the images can be quantified. Because gradient histograms are robust to changes in illumination and minor perturbations, they can effectively avoid interference from high-frequency noise. Compared with traditional pixel difference measurement methods, the second semantic similarity calculation formula has significant advantages. First, it uses gradient histograms to characterize the local structure distribution of an image, thereby improving robustness to changes in illumination and minor perturbations. Second, by processing in blocks and calculating multi-directional gradient histograms, it can capture the edge and texture information of the image more meticulously, thus improving the accuracy of similarity measurement. Finally, since it uses the L1 norm as the metric, it can intuitively reflect the degree of difference between the two histograms in various directions, providing strong support for subsequent image analysis and processing. In summary, the second semantic similarity calculation formula quantifies the similarity between background images by utilizing the difference between gradient histograms and the L1 norm, and has advantages such as robustness to changes in illumination and minor perturbations and avoidance of high-frequency noise interference.

[0039] In a preferred embodiment, the formula for calculating distribution similarity is:

[0040] Where G represents the number of grids into which the background image is divided, and P old (g) represents the frequency with which the target center falls on grid g in the old task, P new (g) represents the frequency with which the target center falls on grid g in the new task, S old (g) Characterizes the ratio of the target area to the grid area in the old task, S new (g) Characterizes the ratio of the target area to the grid area in the new task.

[0041] Furthermore, when dealing with target detection problems in both new and old tasks, the spatial distribution of targets in the background and their size adaptability often differ. Traditional similarity measures, such as Euclidean distance or Wasserstein distance, while able to measure the overall differences between distributions, often fail to accurately capture these specific spatial distribution preferences and size adaptability. Therefore, a distribution similarity calculation formula is designed to more comprehensively evaluate the target distribution similarity between new and old tasks by simultaneously considering the spatial co-occurrence ratio and size matching ratio of targets. Specifically, the distribution similarity calculation formula is as follows: In the distribution similarity calculation formula, the background image is first divided into grids, which are used to capture the spatial distribution information of the target. Then, the frequency of the target center falling in each grid in the old task and the new task is calculated separately, denoted as P. old (g) and P new (g) These frequencies reflect the spatial co-occurrence of the target in different grids. Furthermore, the ratio of the target area to the grid area, denoted as S, is also considered. old (g) and S new (g) measures the target's size fit in each grid. These ratios reflect the degree of size matching of the target across different grids. Finally, the final distribution similarity measure is obtained by multiplying the spatial co-occurrence ratio (i.e., the ratio of frequencies) and the size matching ratio (i.e., some function of the area ratio, such as a direct ratio or a more complex function) for each grid and averaging these products globally. The physical meaning of this formula is that it provides a comprehensive similarity measure for each grid by combining the spatial co-occurrence ratio and the size matching ratio. The spatial co-occurrence ratio reflects the target's spatial distribution preference across different grids, while the size matching ratio measures the target's size fit in each grid. By multiplying these two ratios and averaging them globally, a measure that comprehensively reflects the target distribution similarity between new and old tasks is obtained. Compared with traditional similarity measurement methods, it does not require the use of complex distance metrics such as Wasserstein distance, thus reducing computational complexity and improving computational efficiency. Secondly, since it considers both the target's spatial distribution preference and size fit simultaneously, it can provide a more comprehensive and accurate similarity measure. Furthermore, the similarity of each grid can be directly derived from the product of the spatial co-occurrence ratio and the size matching ratio, making the results easier to understand and analyze. In summary, the distribution similarity calculation formula provides an efficient, accurate, and highly interpretable similarity measurement method by simultaneously modeling the spatial distribution preference and size adaptability of targets in the background.

[0042] In a preferred embodiment, task similarity analysis is performed based on the first semantic similarity, the second semantic similarity, and the distributional similarity to obtain task similarity, including:

[0043] The process involves: summing the similarities of the first semantic similarity, the second semantic similarity, and the distribution similarity to obtain a summed similarity value; iterating through the first semantic similarity, the second semantic similarity, and the distribution similarity, and comparing each with the summed similarity value to obtain a first weight, a second weight, and a distribution weight; processing the first weight, the second weight, and the distribution weight using a weighted distribution model to output a fused weighted distribution result; and performing task similarity weighted calculation on the first semantic similarity, the second semantic similarity, and the distribution similarity to obtain the task similarity. The weighted distribution model is generated through training on a decision tree using multiple sets of data. Each set of data includes a weighted array recording data and labels identifying the weighted distribution. The labels for the weighted distribution are identified using the DeFell method. The element positions of the weighted array correspond one-to-one with the first semantic similarity, the second semantic similarity, and the distribution similarity.

[0044] Specifically, to ensure the final task similarity is both comprehensive and accurate, while avoiding over-reliance on similarity parameters with small differences, the first semantic similarity, second semantic similarity, and distributional similarity are summed to obtain a summed similarity value, which represents the overall contribution of all similarity measures. Then, these three similarities are iterated over, compared individually with the summed similarity value, and their respective weights are calculated. These weights reflect the relative importance of each similarity in the overall similarity. This step reveals which similarities dominate the task similarity analysis and which are relatively less important. Simultaneously, to avoid over-reliance on similarity parameters with small differences, a weight distribution model is used to dynamically adjust these weights. This model is generated based on multiple sets of data trained using decision trees and can automatically adjust the weight distribution according to different weight arrays. Each set of these training data includes a weight array recording the data and corresponding labels identifying the weight distribution. These labels are obtained through the DeFell method (an expert survey method), ensuring the accuracy and rationality of the weight distribution. Finally, the weighted distribution results processed by the weighted distribution model are used to perform a weighted calculation on the first semantic similarity, the second semantic similarity, and the distributional similarity. This step comprehensively considers the contribution of all similarities and dynamically adjusts them according to their relative importance, thus obtaining the final task similarity. This method of dynamically assigning weights ensures that task similarity analysis does not overemphasize similarity parameters with small differences. Instead, it allocates weights based on the actual contribution of each similarity, making the final task similarity more accurate and comprehensive. This method has significant advantages in dealing with complex task similarity problems and can provide a more reliable basis for decision-making. In summary, by integrating multiple similarity measures, calculating weights, adopting a dynamically weighted distribution model, and performing weighted calculations, a comprehensive and accurate task similarity can be obtained. This process not only improves the accuracy of task similarity analysis but also provides a flexible and interpretable method to understand and evaluate the similarity between tasks.

[0045] The incremental learning-supported computer vision target recognition method provided in this embodiment of the invention has at least the following technical effects:

[0046] 1. By calculating the first semantic similarity (similarity between target images), the second semantic similarity (similarity between high-frequency background images), and the distribution similarity (similarity of target distribution features in the background), a multi-dimensional comprehensive evaluation of task similarity is achieved. This method not only considers the similarity of the targets themselves but also incorporates background information and target distribution features, thereby improving the accuracy and comprehensiveness of task similarity evaluation.

[0047] 2. High-frequency background images are intelligently selected from historical background images using semantic similarity clustering. These images represent common background environments in the task. This method effectively reduces background image redundancy while retaining key background information, providing efficient and targeted data support for subsequent similarity calculations and incremental model learning.

[0048] 3. By employing a dynamic weight allocation method, the weights of each similarity level in the final task similarity calculation are automatically adjusted based on their respective proportions and the weight distribution model trained through decision trees. Furthermore, depending on the magnitude of task similarity, a shared backbone network incremental learning strategy or an additional branch network incremental learning strategy is flexibly selected to adapt to tasks with different similarity levels. This improves the model's adaptability and learning efficiency, ensuring the flexibility and accuracy of target recognition in dynamic environments.

[0049] Example 2:

[0050] like Figure 2 As shown, based on the same inventive concept as the incremental learning-supported computer vision target recognition method provided in Embodiment 1, this embodiment of the invention also provides an incremental learning-supported computer vision target recognition system, including:

[0051] The image recognition module 11 is used to obtain the associated target image and the first high-frequency background image of the basic target detection model, as well as the target image to be recognized and the second high-frequency background image.

[0052] The first similarity calculation module 12 is used to calculate the first semantic similarity between the associated target image and the target image to be identified.

[0053] The second similarity calculation module 13 is used to calculate the second semantic similarity between the first high-frequency background image and the second high-frequency background image.

[0054] The distribution similarity calculation module 14 is used to calculate the distribution similarity between the first distribution feature of the associated target image in the first high-frequency background image and the second distribution feature of the target image to be identified in the second high-frequency background image.

[0055] The task analysis module 15 is used to perform task similarity analysis based on the first semantic similarity, the second semantic similarity, and the distribution similarity to obtain task similarity.

[0056] The target recognition module 16 is used to perform target recognition by incremental learning of the basic target detection model using a shared backbone network when the task similarity is greater than or equal to the task similarity threshold; otherwise, it performs target recognition by incremental learning of the basic target detection model using an added branch network.

[0057] Furthermore, the image recognition module 11 is also used to perform the following steps:

[0058] Obtain the first historical background image up to the Nth historical background image of the associated target image; perform semantic similarity clustering on the first historical background image up to the Nth historical background image to obtain multiple groups of historical background images; randomly select one image from each class whose number of images in the group is greater than or equal to the threshold of the number of images in the group and add it to the first high-frequency background image.

[0059] Furthermore, the image recognition module 11 is also used to perform the following steps:

[0060] Perform pairwise semantic similarity calculations on the first historical background image up to the Nth historical background image to obtain several semantic similarities; extract k semantic similarities from the several semantic similarities that are closest to the first historical background image, set as the first adjacent semantic similarity, and calculate the mean to obtain the first distribution density; continue until the Nth distribution density is obtained; calculate the mean density of the first distribution density up to the Nth distribution density, and compare it with the first distribution density up to the Nth distribution density to obtain the first outlier factor up to the Nth outlier factor; based on the outlier factor threshold, and combined with the first outlier factor up to the Nth outlier factor, cluster the first historical background image up to the Nth historical background image to obtain the multiple groups of historical background images.

[0061] Furthermore, the first similarity calculation module 12 also includes:

[0062] The formula for calculating the first semantic similarity is: in, The feature covariance matrix of the target image X at the k-th layer of the feature extraction network. The feature covariance matrix ω of the target image Y in the k-th layer of the feature extraction network. k The weight coefficients represent the features of the k-th layer, K represents the number of feature layers selected, Tr(·) represents the matrix trace operation, and represents the energy of the covariance matrix.

[0063] Furthermore, the second similarity calculation module 13 also includes:

[0064] The formula for calculating the second semantic similarity is: Among them, I bg1 The first high-frequency background image is characterized by extraction using a Butterworth high-pass filter. bg2The second high-frequency background image is represented by Butterworth high-pass filtering. H(·) represents the multi-directional gradient histogram of the image block. The image is divided into B blocks, and an 8-directional gradient histogram is calculated for each block. ||·||1 represents the L1 norm and measures the histogram difference.

[0065] Furthermore, the distribution similarity calculation module 14 also includes:

[0066] The formula for calculating distribution similarity is: Where G represents the number of grids into which the background image is divided, and P old (g) represents the frequency with which the target center falls on grid g in the old task, P new (g) represents the frequency with which the target center falls on grid g in the new task, S old (g) Characterizes the ratio of the target area to the grid area in the old task, S new (g) Characterizes the ratio of the target area to the grid area in the new task.

[0067] Furthermore, the task analysis module 15 is also used to perform the following steps:

[0068] The process involves: summing the similarities of the first semantic similarity, the second semantic similarity, and the distribution similarity to obtain a summed similarity value; iterating through the first semantic similarity, the second semantic similarity, and the distribution similarity, and comparing each with the summed similarity value to obtain a first weight, a second weight, and a distribution weight; processing the first weight, the second weight, and the distribution weight using a weighted distribution model to output a fused weighted distribution result; and performing task similarity weighted calculation on the first semantic similarity, the second semantic similarity, and the distribution similarity to obtain the task similarity. The weighted distribution model is generated through training on a decision tree using multiple sets of data. Each set of data includes a weighted array recording data and labels identifying the weighted distribution. The labels for the weighted distribution are identified using the DeFell method. The element positions of the weighted array correspond one-to-one with the first semantic similarity, the second semantic similarity, and the distribution similarity.

[0069] Through the foregoing detailed description of the incremental learning-supported computer vision target recognition method, those skilled in the art can clearly understand the incremental learning-supported computer vision target recognition system in this embodiment. As the system disclosed in the embodiment corresponds to the method disclosed in the embodiment, the description is relatively simple, and relevant parts can be referred to the method section description.

[0070] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A computer vision target recognition method supported by incremental learning, characterized in that, include: Obtain the associated target image and the first high-frequency background image of the basic target detection model, as well as the target image to be identified and the second high-frequency background image, including: Obtain the first historical background image up to the Nth historical background image of the associated target image; Semantic similarity clustering is performed on the first historical background image up to the Nth historical background image to obtain multiple groups of historical background images. For each group whose number of images is greater than or equal to the threshold of the number of images in the group, one image is randomly selected and added to the first high-frequency background image. The first semantic similarity between the associated target image and the target image to be identified is calculated using the following formula: ； in, The feature covariance matrix of the target image X at the k-th layer of the feature extraction network. The feature covariance matrix representing the target image Y at the k-th layer of the feature extraction network. The weight coefficients representing the features of the k-th layer, The number of feature layers selected represents the number of feature layers. The trace operation represents the energy of the covariance matrix; The second semantic similarity between the first high-frequency background image and the second high-frequency background image is calculated using the following formula: ； in, The first high-frequency background image is characterized by extraction using a Butterworth high-pass filter. The second high-frequency background image is characterized by extraction using a Butterworth high-pass filter. To characterize the multi-directional gradient histogram of an image patch, the image is divided into B patches, and an 8-directional gradient histogram is calculated for each patch. Characterizes the L1 norm and measures histogram differences; Calculate the distribution similarity between the first distribution feature of the associated target image in the first high-frequency background image and the second distribution feature of the target image to be identified in the second high-frequency background image. The distribution similarity calculation formula is as follows: ； in, The number of grids that represent the background image is divided into. In the old task, the target center falls on the grid. frequency, In the new task, the target center falls within the grid. frequency, Characterizes the ratio of the target area to the grid area in the old task. Characterizes the ratio of the target area to the grid area in the new task; Task similarity is obtained by performing task similarity analysis based on the first semantic similarity, the second semantic similarity, and the distribution similarity. When the task similarity is greater than or equal to the task similarity threshold, the basic target detection model is subjected to incremental learning of a shared backbone network for target recognition; otherwise, the basic target detection model is subjected to incremental learning of an additional branch network for target recognition.

2. The method as described in claim 1, characterized in that, Semantic similarity clustering is performed on the first historical background image up to the Nth historical background image to obtain multiple sets of historical background images, including: Perform pairwise semantic similarity calculations on the first historical background image up to the Nth historical background image to obtain several semantic similarity scores; Extract k semantic similarities from the first historical background image from near to far from the aforementioned semantic similarities, set them as the first adjacent semantic similarity, and calculate the mean to obtain the first distribution density; Until the Nth distribution density is obtained; Calculate the mean density from the first distribution density to the Nth distribution density, and compare it with the first distribution density to the Nth distribution density respectively to obtain the first outlier to the Nth outlier. Based on the outlier threshold, and combined with the first outlier up to the Nth outlier, the first historical background image up to the Nth historical background image are clustered to obtain the multiple groups of historical background images.

3. The method as described in claim 1, characterized in that, Task similarity analysis is performed based on the first semantic similarity, the second semantic similarity, and the distributional similarity to obtain task similarity, including: Obtain the sum of the similarities of the first semantic similarity, the second semantic similarity, and the distribution similarity to obtain the similarity sum value; Iterate through the first semantic similarity, the second semantic similarity, and the distribution similarity, and compare each with the sum of the similarities to obtain the first weight, the second weight, and the distribution weight; The first weight, the second weight, and the distribution weight are processed using a weight distribution model to output a fused weight distribution result. The first semantic similarity, the second semantic similarity, and the distribution similarity are then weighted by task similarity calculation to obtain the task similarity. The weight distribution model is generated by training multiple sets of data based on a decision tree. Each set of data includes a weight array that records data and labels that identify the weight distribution. The labels of the weight distribution are identified using the DeFell method. The positions of the elements in the weight array correspond one-to-one with the first semantic similarity, the second semantic similarity, and the distribution similarity.

4. A computer vision target recognition system supported by incremental learning, characterized in that, A computer vision object recognition method supported by incremental learning as described in any one of claims 1-3, comprising: The image recognition module is used to obtain the associated target image and the first high-frequency background image of the basic target detection model, as well as the target image to be recognized and the second high-frequency background image, including: Obtain the first historical background image up to the Nth historical background image of the associated target image; Semantic similarity clustering is performed on the first historical background image up to the Nth historical background image to obtain multiple groups of historical background images. For each group whose number of images is greater than or equal to the threshold of the number of images in the group, one image is randomly selected and added to the first high-frequency background image. The first similarity calculation module is used to calculate the first semantic similarity between the associated target image and the target image to be identified. The formula for calculating the first semantic similarity is: ； in, The feature covariance matrix of the target image X at the k-th layer of the feature extraction network. The feature covariance matrix representing the target image Y at the k-th layer of the feature extraction network. The weight coefficients representing the features of the k-th layer, The number of feature layers selected represents the number of feature layers. The trace operation represents the energy of the covariance matrix; The second similarity calculation module is used to calculate the second semantic similarity between the first high-frequency background image and the second high-frequency background image. The formula for calculating the second semantic similarity is: ； in, The first high-frequency background image is characterized by extraction using a Butterworth high-pass filter. The second high-frequency background image is characterized by extraction using a Butterworth high-pass filter. To characterize the multi-directional gradient histogram of an image patch, the image is divided into B patches, and an 8-directional gradient histogram is calculated for each patch. Characterizes the L1 norm and measures histogram differences; The distribution similarity calculation module is used to calculate the distribution similarity between the first distribution feature of the associated target image in the first high-frequency background image and the second distribution feature of the target image to be identified in the second high-frequency background image. The distribution similarity calculation formula is as follows: ； in, The number of grids that represent the background image is divided into. In the old task, the target center falls on the grid. frequency, In the new task, the target center falls within the grid. frequency, Characterizes the ratio of the target area to the grid area in the old task. Characterizes the ratio of the target area to the grid area in the new task; The task analysis module is used to perform task similarity analysis based on the first semantic similarity, the second semantic similarity, and the distribution similarity to obtain task similarity. The target recognition module is used to perform target recognition by incrementally learning a shared backbone network on the basic target detection model when the task similarity is greater than or equal to the task similarity threshold; otherwise, it performs target recognition by incrementally learning a branch network on the basic target detection model.

Images