A vehicle image recognition system based on local feature learning
The vehicle image recognition system based on local feature learning solves the problems of complex background interference and intra-class differences in vehicle recognition in static images by dynamically selecting a local classifier through superpixel segmentation and local feature clustering, thus achieving high-precision and stable vehicle recognition.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Filing Date
- 2026-03-20
- Publication Date
- 2026-07-14
AI Technical Summary
Existing deep learning-based vehicle recognition models suffer from problems such as insufficient local feature extraction in static images, insufficient robustness to intra-class differences, severe interference from complex backgrounds, and difficulty in distinguishing vehicles with similar appearances, resulting in insufficient recognition accuracy and stability.
A vehicle image recognition system based on local feature learning is adopted. Through superpixel segmentation, dense feature extraction, local gradient feature clustering, and local learning model construction, the most relevant local classifier is dynamically selected for decision-making. Combined with incremental learning capabilities, feature representation and classification decision are optimized.
It significantly improves the accuracy and stability of vehicle recognition, can efficiently identify vehicle targets in complex scenes, adapts to changes in lighting, viewing angle and vehicle type, reduces background interference, and has continuous optimization capabilities.
Smart Images
Figure CN122392002A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of vehicle image recognition, and in particular to a vehicle image recognition system based on local feature learning. Background Technology
[0002] With the rapid development of intelligent transportation systems and autonomous driving technology, vehicle image recognition has become one of the key technologies in the field of environmental perception. This technology is widely used in scenarios such as traffic monitoring, violation evidence collection, intelligent parking, and assisted driving. Its core objective is to accurately and quickly detect and identify vehicle targets in complex road environments. Traditional methods largely rely on video-based sequence analysis, which can utilize background subtraction to extract foreground moving regions, thereby suppressing background interference to some extent. However, in modern applications, with the increasing prevalence of traffic reporting using static images captured by mobile devices and the rising real-time requirements of in-vehicle vision systems, the demand for vehicle recognition based on static images is becoming increasingly prominent. Unlike video sequences, static images cannot utilize temporal information for background modeling; all discrimination information originates from single-frame images. This not only makes background interference more pronounced but also introduces significant intra-class differences due to factors such as shooting angle, lighting conditions, vehicle model, color, and occlusion, greatly increasing the difficulty of accurate recognition.
[0003] While general object detection models based on deep learning have made significant progress in vehicle detection, they still have several limitations when dealing with vehicle recognition in static images. First, these models have limited accuracy in recognizing small, partially occluded, or truncated vehicle targets in images, stemming from insufficient integration of subtle local features and global semantic relationships in their feature extraction mechanisms, making it difficult to accurately focus on discriminative vehicle components in complex backgrounds. Second, most methods rely on global feature representations or a single global model, lacking robustness to intra-class differences caused by variations in vehicle model, color, and viewing angle, failing to effectively utilize local feature learning to address the significant fluctuations in appearance among similar vehicles. Furthermore, in complex environments, sensor imaging quality degrades, making single-visual-modality methods susceptible to interference, while multi-sensor fusion strategies often suffer from excessive algorithmic complexity, making them unsuitable for real-time processing. Finally, existing methods perform poorly in distinguishing vehicles with highly similar appearances, lacking a mechanism for effectively co-modeling local detail features and global structural information, thus limiting the model's performance in fine-grained recognition tasks. Therefore, based on these challenges, this invention proposes a vehicle image recognition system based on local feature learning. Summary of the Invention
[0004] To address the aforementioned issues, the present invention aims to provide a vehicle image recognition system based on local feature learning. This system does not rely on a single global model but instead constructs a recognition framework based on a local learning mechanism to process image content in a more refined manner. This effectively enhances the ability and robustness of identifying vehicle targets under different vehicle models, colors, viewing angles, and lighting conditions, ultimately achieving high-precision and high-stability automatic recognition of vehicle areas in complex scenes.
[0005] To achieve the above objectives, this invention provides a vehicle image recognition system based on local feature learning. The system first performs superpixel segmentation on the input image, decomposing it into multiple visually coherent basic units. Then, within each superpixel, a dense sampling strategy is used to extract scale-invariant local gradient features. Furthermore, by performing cluster analysis on the features of all training images, the scattered feature points within each superpixel are reorganized into a normalized statistical histogram representing its visual content composition pattern, which serves as the feature representation of that superpixel. Crucially, the scheme introduces a local learning model construction mechanism. This mechanism optimally selects multiple local learning centers located near the classification boundary from positive samples and dynamically constructs a class-balanced local training subset around each center, containing neighboring positive and negative samples. Finally, the system trains a dedicated local classifier for each local subset. During the recognition phase, by calculating the similarity between the test superpixel features and each local learning center, the system adaptively selects the most relevant local classifier for decision-making, and integrates the discrimination results of all superpixels to accurately delineate the vehicle region.
[0006] In a first aspect, the present invention provides a vehicle image recognition system based on local feature learning, comprising: The image preprocessing module is used to denoise the input image; The image segmentation module is used to segment the preprocessed image into multiple visually coherent superpixel regions; The dense feature extraction module is used to extract local gradient features of the image within the superpixel region using a dense sampling strategy, and to construct a statistical histogram based on the gradient direction. The feature recombination and representation module is used to perform cluster analysis on the local gradient features of all superpixels, and reconstruct a normalized class distribution histogram based on the clustering results to serve as the final feature representation of the superpixel. The local learning model construction module is used to select representative samples from the positive sample space as local learning centers based on superpixel feature representation, and dynamically construct a local training subset composed of neighboring positive and negative samples around each center; The collaborative classification module includes multiple local classifiers that correspond one-to-one with each local training subset. These classifiers are trained separately and, during the recognition phase, adaptively select the most relevant local classifier for category discrimination by calculating the feature similarity between the test sample and each local learning center. Finally, the discrimination results of all superpixels are combined to determine the vehicle region.
[0007] Furthermore, the image segmentation module is configured to perform multi-scale superpixel segmentation, or to optimize the segmentation granularity to ensure that the vast majority of superpixel regions spatially belong to a single semantic category in the vehicle target or background. This process reduces the number of superpixels coexisting with the background from the source, laying a more accurate unit foundation for subsequent feature extraction and classification, thereby reducing the impact of background interference on the recognition effect.
[0008] Furthermore, the dense feature extraction module performs dense sampling on all pixels or pixels with a specified step size in the image through a sliding window, calculates the gradient information of pixels in the local neighborhood of each sampling point, quantizes the gradient direction into multiple intervals to form a direction histogram, and concatenates and combines the direction histograms of multiple local regions into a multi-dimensional feature descriptor. This process ensures the completeness of feature description by covering all significant local structures, thereby enhancing the ability to identify subtle parts of the vehicle.
[0009] Furthermore, the feature recombination and representation module includes: The feature clustering unit is used to perform unsupervised clustering of local gradient features extracted from all training images to obtain typical patterns in the feature space. The distribution statistics unit is used to calculate the distribution of cluster categories of all local features within each superpixel. Feature normalization unit is used to transform distribution statistics results into normalized statistical histograms.
[0010] Furthermore, the category distribution histogram obtained by clustering local gradient features in the feature recombination and representation module has the same dimension as the number of categories formed by clustering, and the distribution histogram is used to characterize the visual content composition pattern of the corresponding superpixel. This representation method aggregates local features into more discriminative statistical features, thereby providing a higher level of semantic description while retaining detailed information.
[0011] Furthermore, the local learning center selection strategy in the local learning model construction module includes: The initial seed point generation unit is used to randomly select several samples from the positive sample set as initial center points; The center point optimization unit is used to iteratively calculate and move each initial center point sequentially towards its nearest negative sample and then towards its nearest positive sample to determine the final optimized center point located near the decision boundary between positive and negative samples.
[0012] Furthermore, the local learning model construction module also includes a local sample set expansion unit, which is used to select a certain number of samples that are closest to each other from the positive sample set and the negative sample set in the feature space, with each optimized center point as the core, to jointly form a locally balanced training subset.
[0013] Furthermore, during the training phase, the collaborative classification module independently trains a corresponding binary classifier for each local training subset, forming a set of local classifiers. During the prediction phase, for each superpixel feature to be identified, the local classifier corresponding to the center with the smallest distance is selected to make a classification decision by calculating its feature distance to each local learning center. This mechanism can call the most matching classification model according to the specific characteristics of the test sample, thereby improving recognition accuracy and scene adaptability.
[0014] Furthermore, the collaborative classification decision module is configured as a dynamic weighted fusion mechanism, which comprehensively calculates the feature similarity between the test sample and each local learning center, as well as the discrimination confidence of each local classifier for the sample, assigns appropriate weights to the outputs of all local classifiers, and makes weighted decisions accordingly, rather than relying solely on the results of a single nearest neighbor classifier.
[0015] Furthermore, the system also includes an adaptive superpixel segmentation module, which dynamically determines the segmentation granularity based on the visual complexity and boundary sharpness of the image region, and automatically determines whether to re-segment the region through a stability evaluation function.
[0016] Furthermore, the system is also equipped with an incremental learning management module, which provides online update capability for the local classifier in the local learning model construction module. This module can collect high-value samples encountered during the recognition process and trigger parameter updates of the corresponding local classifier when the system is idle, thereby enabling the system to continuously optimize and adapt to new scenarios.
[0017] Furthermore, the system employs a statistical distribution similarity-based metric to calculate the distance between superpixel features. This metric assesses the similarity of two feature histograms by calculating the degree of overlap in each dimension. The higher the degree of overlap, the smaller the distance. This method is more suitable for comparing feature representations of different distribution types, thus more accurately reflecting the similarity of visual content during the measurement process.
[0018] Furthermore, the system constructs a recognition network composed of multiple local expert models through the local learning model construction module and the collaborative classification module. This network can adaptively handle intra-class differences in vehicle images caused by changes in model, color, viewing angle, and lighting, and effectively suppress interference from complex backgrounds. Through the collaborative work of distributed local models, this network achieves a more powerful understanding and analysis capability for complex visual scenes.
[0019] This invention provides a vehicle image recognition system based on local feature learning. The system first decomposes the image into semantically coherent units through adaptive superpixel segmentation. Then, it employs a dense feature extraction and recombination strategy to aggregate the local gradient information within each superpixel into a scale- and deformation-robust statistical distribution histogram feature. The system introduces a boundary-aware local learning model construction mechanism. By intelligently selecting key center points in the feature space and constructing multiple class-balanced local training subsets, it refines the modeling of intra-class differences and complex decision boundaries for vehicles. Finally, through a collaborative classification decision module, the system adaptively selects the most relevant local classifier for each test sample, integrating the recognition results of all units to achieve accurate localization of the vehicle region.
[0020] This system, through its local learning and collaborative decision-making architecture, effectively overcomes the challenge of intra-class discrepancies caused by variations in vehicle model, color, viewing angle, and lighting, while significantly suppressing interference from complex backgrounds. Its feature recombination method enhances robustness to local structural changes, while the collaborative working mode of the distributed local models strengthens the system's generalization ability and discrimination accuracy in complex scenes. Ultimately, this scheme can achieve highly accurate and stable vehicle target recognition from static images without relying on temporal information, providing more reliable technical support for intelligent traffic management and visual perception applications.
[0021] Beneficial effects By implementing the vehicle image recognition system based on local feature learning provided by the present invention, the following technical effects are achieved: (1) Cluster-based feature recombination and representation methods aggregate scattered local features into statistical distribution histograms. This creates a feature representation that is more robust to scale changes and local deformations, and elevates low-level gradient information to high-level semantic features that characterize visual composition patterns, thereby significantly enhancing the classification and discrimination capabilities of feature representation.
[0022] (2) The boundary-aware local learning center selection strategy optimizes the local learning center to be near the decision boundary of positive and negative samples. This ensures that each local model can accurately focus on the feature space region with the highest classification difficulty, thereby greatly improving the learning ability of the local model set to complex decision boundaries and the discrimination precision of the overall model.
[0023] (3) The adaptive superpixel segmentation method achieves high-precision alignment between superpixel regions and local semantic content of the image. It significantly improves the semantic purity of superpixel units from the source, effectively reduces the number of mixed superpixels where vehicles and backgrounds coexist, thus laying a more reliable foundation for subsequent feature extraction and classification, and fundamentally reducing the introduction of background interference.
[0024] (4) The dynamic weight allocation collaborative decision-making mechanism innovates the single decision-making mode by integrating the weighted opinions of multiple local classifiers. It significantly enhances the system's robustness in discriminating against difficult samples with abnormal appearance, partial occlusion, or complex backgrounds. By comprehensively considering feature similarity and classifier confidence, it effectively avoids recognition failures caused by misjudgment by a single local model, thereby improving the reliability of the overall decision.
[0025] (5) Incremental online learning capability endows the system with continuous self-optimization characteristics. It enables the system to actively absorb new knowledge from real application scenarios and correct inherent errors, thereby achieving rapid adaptation to new vehicle targets and autonomous correction of historical misjudgment patterns without global reconstruction, ensuring the performance maintenance and evolution of the system in long-term deployment. Attached Figure Description
[0026] To make the above-described vehicle image recognition system based on local feature learning of the present invention more apparent and understandable, the accompanying drawings used in the specific embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0027] Figure 1 This represents a schematic diagram of the local learning structure. Figure 2 A schematic diagram showing the segmentation of a superpixel region with a size of rs=10; Figure 3 A schematic diagram showing the segmentation of a superpixel region with a size of rs=50; Figure 4 This is a schematic diagram showing the segmentation of a superpixel region with a size of rs=100. Detailed Implementation
[0028] Example 1: This embodiment provides a vehicle image recognition method based on local learning. The method first preprocesses the input image, using Gaussian smoothing filtering to remove noise interference introduced during image acquisition, providing optimized image data for subsequent processing. Then, superpixel segmentation is performed on the preprocessed image, dividing it into multiple visually coherent regional units. During superpixel segmentation, it is necessary to balance the segmentation granularity. Overly fine segmentation leads to overly localized features lacking semantic information, while overly coarse segmentation causes a single superpixel to simultaneously contain vehicle and background information, affecting the purity of subsequent feature representation. Experiments determine appropriate segmentation parameters so that the vast majority of superpixels can be completely assigned to a single semantic category within either the vehicle or the background, laying a solid foundation for subsequent recognition.
[0029] In the feature extraction stage, the Dense Scale Invariant Feature Transform (DSIFT) method is employed, using superpixels as the basic unit for feature description. Unlike the traditional SIFT method, which relies on keypoint detection, DSIFT uses a dense sampling strategy. It traverses all pixels in the image with a sliding window at unit steps, calculating the gradient information of its local neighborhood at each sampling point. Specifically, for each sampling point, the gradient magnitude and direction within its 4×4 pixel region are calculated. The gradient direction is quantized into eight directional intervals, forming a local gradient direction histogram. Finally, the gradient histograms of all sub-regions are concatenated to construct a 128-dimensional feature descriptor. This dense sampling method ensures that sufficient feature points are obtained within each superpixel, thus effectively describing the visual characteristics of the superpixel.
[0030] After obtaining the DSIFT features of all superpixels, the features need to be reorganized and represented. The DSIFT features extracted from all images in the training set are combined into a large feature set. The k-means clustering algorithm is then used to divide these features into k categories, resulting in k cluster centers. For each superpixel, the distribution of all DSIFT feature points within these cluster categories is statistically analyzed. A statistical histogram is constructed with the cluster category as the x-axis and the number of feature points in the corresponding category as the y-axis. Since different superpixels may contain different numbers of feature points, the histogram is normalized to eliminate this influence by dividing each y-axis value by the total number of feature points in that superpixel. The resulting normalized statistical histogram is the final feature representation of the superpixel. This feature representation method aggregates scattered local features into statistically distributed features with stronger discriminative power.
[0031] During the model training phase, a local learning strategy is employed to construct the recognition model. The local learning structure is as follows: Figure 1As shown, firstly, based on the ground truth information of the training samples, the superpixel feature samples are divided into a positive sample set and a negative sample set, corresponding to the vehicle target and the background, respectively. Then, a local learning center selection strategy is used to determine multiple local learning centers. This strategy randomly selects several samples from the positive sample set as initial seed points. Through iterative calculation, each seed point is moved sequentially towards its nearest negative neighbor, and then towards its nearest positive neighbor, ultimately obtaining key sample points near the decision boundary between positive and negative samples as local learning centers. Around each local learning center, a certain number of nearest positive and negative samples are selected in the feature space to form a locally balanced training subset. Finally, a support vector machine classifier is trained for each local training subset, forming a classifier set composed of multiple local classifiers.
[0032] In the prediction phase, the test image is also preprocessed, superpixel segmented, and DSIFT feature extracted. Statistical histogram features for each superpixel are generated in the same manner as in the training phase. The superpixel segmentation is as follows: Figure 2-4 As shown in the diagram. For each superpixel feature to be identified, its Bhattacharyya distance to each local learning center obtained during the training phase is calculated. The SVM classifier corresponding to the nearest local learning center is selected for class discrimination. Bhattacharyya distance evaluates similarity by calculating the degree of overlap between two histogram features in each dimension; the higher the overlap, the smaller the distance. This distance metric is particularly suitable for comparing feature representations of different distribution types. Finally, all superpixel regions identified as vehicle targets are merged to obtain the identified vehicle region. This method effectively solves the problem of intra-class differences in vehicle images caused by factors such as model, color, and viewing angle through a local learning strategy. At the same time, it reduces background interference through superpixel segmentation and statistical histogram features, achieving good results in static image vehicle recognition tasks.
[0033] Example 2: Building upon the aforementioned embodiments, a vehicle image recognition method based on local learning is further described. In the image preprocessing stage, in addition to conventional Gaussian smoothing filtering for noise reduction, this method can incorporate image enhancement, contrast adjustment, and other preprocessing operations to improve image quality, depending on the specific application scenario. In the superpixel segmentation stage, an adaptive segmentation strategy is employed, dynamically adjusting segmentation parameters based on the texture complexity and boundary strength of local image regions. A finer segmentation granularity is used for regions with complex textures and clear boundaries, while a coarser granularity is used for regions with smooth textures and blurred boundaries. This adaptive segmentation approach can reduce unnecessary computational overhead while preserving important details.
[0034] In the feature extraction and representation stage, in addition to basic gradient information, auxiliary features such as color and texture can be introduced to construct multimodal feature descriptors. During feature recombination, more advanced clustering algorithms, such as hierarchical clustering or spectral clustering, are employed to obtain more representative feature patterns. Furthermore, when constructing superpixel statistical histograms, not only is the category distribution of feature points considered, but spatial distribution information can also be incorporated. For example, superpixels can be divided into several sub-regions and their feature distributions can be statistically analyzed separately, thereby better preserving spatial structure information.
[0035] In the selection of local learning centers, in addition to considering distance relationships in the feature space, confidence weights for samples are introduced, prioritizing samples from regions with high classification uncertainty as local learning centers. When constructing local training subsets, a dynamic sample selection mechanism is employed, adaptively adjusting the size and positive / negative sample ratio of each local training subset based on the characteristics of the test samples, making the local model more targeted. Furthermore, a sample weighting mechanism is introduced; when training the local SVM classifier, appropriate weights are assigned to training samples based on their similarity to the local learning centers, with samples exhibiting higher similarity having a greater influence during training.
[0036] In the classification decision-making stage, instead of simply assigning test samples to the nearest local classifier, the final decision is made by comprehensively considering the discrimination results of multiple relevant local classifiers through weighted voting or stacked generalization. Specifically, for each test sample, its similarity to all local learning centers is calculated, and appropriate weights are assigned to each local classifier based on the similarity. Then, the outputs of each classifier are fused according to their weights to obtain the final classification result. This ensemble decision-making mechanism can effectively reduce the risk of misclassification by a single classifier and improve the robustness of the system.
[0037] Furthermore, an online learning mechanism is introduced, enabling the system to adapt to environmental changes. During system operation, new training samples are continuously collected, including correct and incorrect recognition results from users. When enough new samples are accumulated, a model update process is triggered, incrementally training affected local classifiers, or, in some cases, adding new local learning centers and corresponding classifiers. This online learning capability allows the system to gradually optimize performance and adapt to new vehicle types and environmental conditions.
[0038] In terms of computational efficiency, the system employs multiple acceleration strategies. During feature extraction, integral image technology is used to accelerate gradient calculation; during clustering, approximate nearest neighbor search is used to improve computational speed; and during classification decisions, a search index for local learning centers is established to quickly locate relevant classifiers. These optimizations enable the method to maintain high accuracy while meeting the speed requirements of practical applications.
[0039] Example 3: Building upon the aforementioned embodiments, and addressing the issue of insufficient adaptability of fixed-parameter superpixel segmentation in complex scenes, this method abandons the traditional approach of pre-setting a fixed number of superpixels. Instead, it dynamically determines the optimal segmentation granularity for each region based on the image's inherent information complexity and spatial continuity. By constructing a stability function that integrates local image entropy and boundary intensity information, when the function value falls below an adaptive threshold, the current region is considered to have reached an ideal state of internal uniformity and clear boundaries, thus automatically terminating the segmentation process. This method ensures meticulous segmentation in textured background regions while performing coarse-grained merging on smooth vehicle surfaces, optimizing the unit quality of subsequent feature extraction from the source.
[0040] Input the image to be segmented, and set the initial segmentation block size and stability threshold coefficient.
[0041] For each segmentation block in the current image : In block On a grayscale image, the information entropy of its grayscale histogram is calculated to quantify the texture complexity of the region; Using boundary detection operators such as Sobel, compute blocks The gradient magnitude of all pixels within the block is taken as the average value as the boundary strength of the block. The stability of this region can be calculated using the following formula:
[0042] In the formula, The value represents the stability of the i-th image region. The smaller the value, the more uneven the region is or the less clear its boundaries are, and the more it needs to be segmented again. Let be the local information entropy of the i-th image region. It is a measure of the texture complexity and information content within that region. The higher the entropy value, the more complex the texture and the richer the information contained in that region. It is a balancing weight coefficient, a positive real number, used to adjust the weight of boundary strength in stability calculation, relative to local information entropy; Let be the boundary intensity of the i-th image region. It quantifies the visual contrast between the region and its surrounding areas by calculating the average gradient magnitude of the pixels within the region. The higher the boundary intensity, the clearer the edge of the region.
[0043] Calculate the adaptive threshold In the formula, This is the stability threshold coefficient; This represents the initial average stability across all regions of the entire image. If... Then the region is considered If the region is unstable, split it into four equal sub-regions; otherwise, leave the region unchanged.
[0044] Repeat the above steps until no more regions need to be split, or the preset maximum number of iterations is reached. The final output is an adaptively generated set of irregular superpixels.
[0045] Testing on the Cityscapes vehicle dataset showed that this method effectively reduced the proportion of mixed vehicle-background superpixels from approximately 15% in the traditional SLIC method to below 5%. This improvement significantly enhanced the purity of subsequent feature extraction, ultimately increasing the overall vehicle recognition accuracy by approximately 3.5 percentage points. The results demonstrate that the adaptive superpixel segmentation method based on information entropy and spatial continuity can significantly optimize the semantic purity of image primitives. By dynamically adjusting the segmentation granularity, this method effectively reduces the number of mixed superpixels simultaneously covering both vehicles and background, thereby mitigating background noise interference in subsequent feature extraction and classification processes from the outset.
[0046] Example 4: Building upon the aforementioned embodiments, a dynamic weight allocation mechanism is proposed to address the limitations of traditional nearest neighbor selection classifier strategies. The principle is to move away from solely entrusting the classification decision of a test sample to a single nearest local classifier. Instead, it comprehensively considers the discrimination results of all local classifiers and assigns appropriate weights to each classifier based on its feature relevance to the test sample and its local confidence. By introducing a weighting function based on the Bach coefficient and classifier confidence, local classifiers with more similar feature distributions to the test sample and higher self-determinism are given greater weight in the final decision, thus achieving a more robust and accurate multi-expert knowledge fusion.
[0047] Testing the feature histogram of superpixels , and all N trained local SVM classifiers and their corresponding local learning center features.
[0048] calculate The Bach coefficients for each local center are normalized to form similarity weights:
[0049] In the formula, Similarity weight; This is the Barthel coefficient.
[0050] For each local SVM classifier, obtain its corresponding... The decision function output value. The magnitude of this absolute value reflects the confidence level of the classifier's judgment. Calculate the confidence weights:
[0051] In the formula, Confidence weights; This is the output value of the decision function.
[0052] The two weights are combined to form the overall voting weight for each classifier:
[0053] In the formula, To take into account the overall voting weight; The fusion coefficient is between 0 and 1, used to balance the importance of similarity and confidence.
[0054] Each classifier pair The initial judgment result is multiplied by its comprehensive weight. All weighted results are summed. If the final sum is positive, it is classified as a positive class; otherwise, it is classified as a negative class.
[0055] On the KITTI dataset, compared to a single nearest neighbor classifier selection strategy, the dynamic weight allocation mechanism improves the F1-Score for vehicle recognition from 91.5% to 94.0%. The improvement is particularly significant for difficult samples where vehicles are partially occluded or located at uncommon viewpoints. This is because the mechanism integrates the opinions of multiple relevant but not nearest-neighbor local models, avoiding misclassification by a single model. The results demonstrate that this mechanism overcomes the limitations of single-model decision-making by fusing the discriminative information of multiple relevant local classifiers and weighting them based on their correlation with the test samples and their own confidence levels, thus constructing a more intelligent ensemble decision-making framework.
[0056] Example 5: Building upon the aforementioned embodiments, the system is equipped with the ability to continuously evolve in response to unknown scenarios. The principle behind this is to assign an independent incremental learning interface to each local classifier. When the system encounters misclassified samples with high confidence during the prediction phase, or actively discovers new, high-value unlabeled samples, these samples and their true labels can be added to the training set of the corresponding local classifier. Subsequently, during idle periods or when a certain sample accumulation threshold is reached, the system triggers rapid incremental training of that local classifier. This allows for targeted correction of misperceptions and absorption of new knowledge without affecting other local models, enabling the entire system to continuously adapt to new environments and vehicle types.
[0057] During the prediction phase, the system records samples that are classified with high confidence by the local classifier, but whose final results do not match those of manual review or high-precision fusion. These samples are marked as erroneous samples and stored.
[0058] For unlabeled samples with low confidence, the system can present them to humans for labeling, or under certain conditions, use the results of the system's overall collaborative decision-making as pseudo-labels and store them as new samples.
[0059] The collected samples are used to calculate the Bach distance to each local learning center based on their features, and then assigned to the incremental learning pool corresponding to one or more local classifiers that are closest to them.
[0060] When the number of samples in the incremental learning pool of any local classifier reaches a preset threshold, or when the system enters an idle computing cycle, the incremental update process of that classifier is triggered.
[0061] The original training set of the local classifier is merged with the new samples in its incremental learning pool, and the model parameters of the local classifier are updated using a fast incremental SVM learning algorithm or a mini-batch retraining method. After the update is complete, the incremental learning pool of the classifier is cleared.
[0062] In a month-long real-world road image test, the system with online learning capabilities, with the same initial performance, improved its recall rate for newly appearing specific vehicle models from 70% to 95% within two weeks. Simultaneously, the system's false detection rate for initially easily confused obstacles decreased by 40% within a week after introducing a small number of corrective samples. This result was obtained in comparison to a fixed model version without online learning capabilities.
Claims
1. A vehicle image recognition system based on local feature learning, characterized in that, include: The image preprocessing module is used to denoise the input image; The image segmentation module is used to segment the preprocessed image into multiple visually coherent superpixel regions; The dense feature extraction module is used to extract local gradient features of the image within the superpixel region using a dense sampling strategy, and to construct a statistical histogram based on the gradient direction. The feature recombination and representation module is used to perform cluster analysis on the local gradient features of all superpixels, and reconstruct a normalized class distribution histogram based on the clustering results to serve as the final feature representation of the superpixel. The local learning model construction module is used to select representative samples from the positive sample space as local learning centers based on superpixel feature representation, and dynamically construct a local training subset composed of neighboring positive and negative samples around each center; The collaborative classification module includes multiple local classifiers that correspond one-to-one with each local training subset. These classifiers are trained separately and, during the recognition phase, adaptively select the most relevant local classifier for category discrimination by calculating the feature similarity between the test sample and each local learning center. Finally, the discrimination results of all superpixels are combined to determine the vehicle region.
2. The system according to claim 1, characterized in that: The image segmentation module is configured to perform multi-scale superpixel segmentation, or to optimize the segmentation granularity to ensure that the vast majority of superpixel regions spatially belong to a single semantic category in the vehicle target or background.
3. The system according to claim 1, characterized in that: The dense feature extraction module performs dense sampling on all pixels or pixels with a specified step size in the image through a sliding window, calculates the gradient information of pixels in the local neighborhood of each sampling point, quantizes the gradient direction into multiple intervals to form a direction histogram, and concatenates and combines the direction histograms of multiple local regions into a multi-dimensional feature descriptor.
4. The system according to claim 1, characterized in that, The feature recombination and representation module includes: The feature clustering unit is used to perform unsupervised clustering of local gradient features extracted from all training images to obtain typical patterns in the feature space. The distribution statistics unit is used to calculate the distribution of cluster categories of all local features within each superpixel. Feature normalization unit is used to transform distribution statistics results into normalized statistical histograms.
5. The system according to claim 4, characterized in that: The category distribution histogram obtained by clustering local gradient features in the feature recombination and representation module has the same dimension as the number of categories formed by clustering, and the distribution histogram is used to characterize the visual content composition pattern of the corresponding superpixel.
6. The system according to claim 1, characterized in that, The local learning center selection strategy in the local learning model construction module includes: The initial seed point generation unit is used to randomly select several samples from the positive sample set as initial center points; The center point optimization unit is used to iteratively calculate and move each initial center point sequentially towards its nearest negative sample and then towards its nearest positive sample to determine the final optimized center point located near the decision boundary between positive and negative samples.
7. The system according to claim 6, characterized in that: The local learning model construction module also includes a local sample set expansion unit, which is used to select a certain number of samples that are closest to each optimized center point in the feature space from the positive sample set and the negative sample set respectively, and together form a locally balanced training subset.
8. The system according to claim 1, characterized in that: During the training phase, the collaborative classification module independently trains a corresponding binary classifier using each local training subset, forming a set of local classifiers. During the prediction phase, for each superpixel feature to be identified, the module calculates its feature distance to each local learning center and selects the local classifier corresponding to the center with the smallest distance to make a classification decision.
9. The system according to claim 1, characterized in that: The system also includes an adaptive superpixel segmentation module, which dynamically determines the segmentation granularity based on the visual complexity and boundary clarity of the image region, and automatically determines whether to re-segment the region through a stability evaluation function.
10. The system according to claim 1, characterized in that: The system employs a statistical distribution similarity metric to calculate the distance between superpixel features. This metric assesses similarity by calculating the degree of overlap between two feature histograms across various dimensions. The higher the degree, the smaller the distance.