Efficient target extraction and high-precision classification and grading system under industrial intelligent network connection scene

Abstract

The invention discloses an efficient target extraction and high-precision classification grading system in an industrial intelligent network connection scene, and relates to the field of industrial intelligent network connection. The system comprises a component A, a component B and a component C. The component A performs label distribution based on self-learning semantic perception, the component B performs optimal single-sample progressive distribution and NMS elimination, and the component C performs rotation rectangle regression based on GWD Loss, so that a 'SemSL + GWD + OSG' integrated technology is provided, three core contradictions of rotation frame positioning precision, classification precision and system efficiency are considered, and the positioning accuracy of the rotation frame is improved. The method is particularly suitable for challenge scenes such as multi-direction, dense, small-target and complex background in an industrial intelligent network connection environment, high-precision real-time detection and classification grading can be achieved under the condition of not depending on heavy hardware, meanwhile, the model structure does not need to be changed at all, the method supports embedding into any target detection model in a modular mode, and the detection efficiency is greatly improved. And the method has remarkable engineering feasibility and industrial value.

Description

Technical Field

[0001] This invention relates to the field of industrial intelligent connected vehicles, specifically to an efficient target extraction and high-precision classification and grading system for industrial intelligent connected vehicle scenarios. Background Technology

[0002] Industrial intelligent connected scenarios (such as intelligent manufacturing, intelligent logistics, intelligent security / inspection, low-altitude unmanned systems, etc.) place demands on image target detection with "high precision + high efficiency + strong robustness" at the same time: targets are densely distributed in multiple directions, have large differences in shape and scale, have blurred appearances, and complex backgrounds, and require edge / embedded devices to operate in real time.

[0003] Traditional horizontal bounding box methods are difficult to accurately describe the location of directed targets. During the training and inference of directed target detection models, conventional one-hot hard label segmentation leads to confusion between the background and front-end labels during the training phase, which seriously reduces the accuracy of target classification and grading. Post-processing based on NMS is extremely time-consuming in dense and rotating scenes, becoming a system bottleneck and seriously affecting the running efficiency in lightweight devices. Summary of the Invention

[0004] The purpose of this invention is to address the above-mentioned shortcomings by proposing an efficient target extraction and high-precision classification and grading system for industrial intelligent connected scenarios that uses a rotating frame positioning method to extract oriented targets.

[0005] The present invention specifically adopts the following technical solution:

[0006] A high-efficiency target extraction and high-precision classification and grading system for industrial intelligent connected vehicle scenarios, comprising components A, B, and C, specifically:

[0007] Component A: Label allocation based on self-learning semantic awareness. The specific implementation steps are as follows:

[0008] (1) Construction of output confidence

[0009] Classification accuracy Directly take the network output class confidence score (if it includes a quality branch, then take the product of the two); regression accuracy. Based on prediction boxes With target box of To perform normalization enhancement (increase variance and strengthen the discrimination of high-quality samples), the formula is as follows:

[0010] ;

[0011] If the objective is optimal If the value is ≥0.5, then the optimal sample regression accuracy is approximately 1, and the accuracy of other samples follows the trend. Reduce rapid decline, based on classification accuracy and regression accuracy The output confidence score is obtained using the following formula:

[0012] ;

[0013] The output confidence score can gradually approximate the semantic perception probability during training, where γ is used to adjust the relative weight ratio of the two branches;

[0014] (2) Multidimensional Gaussian initial value, i.e. (semantic prior) initialization

[0015] Initial values ​​are generated only for samples falling into the target region; in rotated scenarios, the interior is determined by the rotation coordinates.

[0016] ;

[0017] in, and From sample to target The center offset; This is the step size for that layer; This represents the distance from the sample to each side of the target. Control level soft matching; , These represent the spatial and hierarchical variances, respectively, and the samples that fall outside the target. ;

[0018] (3) Semantic probability update, i.e. (self-learning)

[0019] Will and Perform adaptive fusion to obtain semantic confidence. (Real-time approximation of semantic probability distribution):

[0020] ;

[0021] in, The system adaptively adjusts based on the highest output confidence level under the current objective, shifting from initial prior dominance to later model dominance. To fine-tune the index;

[0022] (4) Generation of fine-grained soft tags

[0023] For each target g, the sample set Normalization and nonlinear compression yield the final soft tags. (In the case of a single class):

[0024] ;

[0025] in, The steepness of the control curve is controlled; the normalization process ensures that the maximum label of each target is 1, preserving the continuity and difference of soft labels;

[0026] (5) Training and Loss

[0027] Classification loss:

[0028] ;

[0029] Regression loss, GIoU or IoU loss, regression only weights samples where y > 0.

[0030] ;

[0031] Quality Branch: Labels are acceptable The loss is calculated using binary cross-entropy:

[0032] ;

[0033] Total loss:

[0034] ;

[0035] Where N is the number of positive weighted samples;

[0036] Component B: Optimal single-sample progressive allocation and NMS elimination, the specific implementation steps are as follows:

[0037] (1) Positive sample quality score, initial selection of the best

[0038] The set of positive samples obtained from regular label assignment In the middle, the quality score is constructed by fusing the predicted score with the Gaussian prior. :

[0039] ;

[0040] in The generation method is the same as component A. Initial value from Dominated, later by leading, To measure the overall representativeness of the sample to the target;

[0041] Initial selection of the optimal sample: ;

[0042] (2) Gradual adjustment, consistent with quality output convergence

[0043] Let the quality branch output be... If the difference between the maximum and second-largest values ​​increases, it indicates an improvement in uniqueness, and the value should be... Gradually "pulled" to the maximum Location:

[0044] ;

[0045] in To adjust the intensity; The bigger it is, the more you trust it. Guidance; only in the mid-to-late stages When the value is large, strengthen this item to avoid initial instability;

[0046] (3) Quality layer one-to-one label and loss

[0047] For each target Only set As the only positive sample, its label The labels for all other positive and negative samples are 0;

[0048] The loss corresponding to the quality layer is cross-entropy or BCE: ;

[0049] The classification and regression layers still use the conventional multiple positive sample label allocation to avoid extreme imbalance.

[0050] (4) Lightweight suppression replacing NMS during inference period, three-dimensional cube

[0051] The three-dimensional max-pooling cube suppression process is implemented as follows:

[0052] Constructing the suppressed input score: n≈0.4 (improving focus on high scores);

[0053] A 3D max pooling space × layer FPN is used, with kernel = 3×3×3, stride = 1, and boundary padding = 1. Each layer is upsampled to the same scale and then backsampled.

[0054] Mask generation hard suppression: Otherwise, it is 0;

[0055] Mask generation Gaussian soft suppression: ;

[0056] Output score: Select Top-K as the final result, and do not perform NMS again;

[0057] Component C: Rotated Rectangle Regression Based on GWD Loss

[0058] To improve the positioning accuracy of the rotated bounding box, the existing Gaussian Wasserstein DistanceRegression Loss is used to regress the rotated rectangle. First, a two-dimensional Gaussian distribution is used to model the rotated target, and the specific formula is as follows:

[0059] ;

[0060] Therefore, the Gaussian Wasserstein Distance (GWD) is used to replace the non-differentiable rotational IoU, as shown in the following formula:

[0061] ;

[0062] The Loss value is calculated based on GWD using the following formula:

[0063] ;

[0064] Finally, components A, B, and C were integrated and industrialized.

[0065] Preferably, component A uses purely dynamic positive and negative samples.

[0066] Preferably, component A only uses IoU or confidence weighting to explicitly approximate the true semantic probability.

[0067] The present invention has the following beneficial effects:

[0068] The "SemSL+GWD+OSG" integrated technology provided by this invention balances the three core contradictions of rotation frame positioning accuracy, classification accuracy, and system efficiency. It is particularly suitable for challenging scenarios such as multi-directional, dense, small targets and complex backgrounds in industrial intelligent connected environments. It can achieve high-precision real-time detection and classification without relying on heavy hardware, and at the same time, it does not require any changes to the model structure. It can be embedded into any target detection model in a modular way, and has significant engineering feasibility and industrial value.

[0069] In industrial scenarios, where the computing power of intelligent connected systems is limited, this solution develops an end-to-end directional target extraction and classification technology for industrial intelligent connected systems. It optimizes and innovates high-precision classification and non-maximum suppression, achieving rapid and compact target extraction.

[0070] This solution constructs semantically aware soft labels using a self-learning approach for classification and grading training, significantly improving the accuracy of foreground / class distinction and enhancing adaptability to irregular, ambiguous, and complex background targets.

[0071] This solution achieves optimal single-sample progressive allocation during the training phase and completely eliminates the NMS process during the inference phase through a lightweight suppression strategy without introducing additional complex modules, significantly improving the overall detection speed, especially suitable for the real-time requirements of rotating boxes and edge devices.

[0072] This solution uses the existing Gaussian Wasserstein Distance Regression Loss (GWDLoss) to perform regression on the rotated rectangle, which solves the boundary discontinuity problem in the training process and improves the positioning accuracy.

[0073] The three components can be decoupled and integrated to form an integrated technical solution for efficient target extraction and high-precision classification and grading of industrial intelligent connected systems. It can be effectively deployed on lightweight, embedded platforms and support the implementation of real-time systems. Attached Figure Description

[0074] Figure 1 A visual comparison of the effects of different methods for separating positive and negative samples;

[0075] Figure 2 A schematic diagram of a conventional NMS method;

[0076] Figure 3 Overall structure diagram of the positive and negative sample partitioning method based on self-learning semantic awareness (SemSL);

[0077] Figure 4 Overall structure diagram based on the Optimal Single Sample Progressive Matching (OSG) method;

[0078] Figure 5 Visual diagram of key steps in Component A;

[0079] Figure 6 A visual illustration of the optimal single-sample progressive label assignment;

[0080] Figure 7 Schematic diagram of 3D max pooling principle;

[0081] Figure 8 GWD distance diagram. Detailed Implementation

[0082] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings and specific examples:

[0083] Intelligent Connected Systems: Intelligent connected systems are systems built upon the internet and communication technologies to connect devices and services for communication and interaction. Through technologies such as sensors, cloud computing, and artificial intelligence, intelligent connected systems achieve real-time data transmission, status monitoring, and control optimization across different terminal devices, and are widely used in smart homes, autonomous driving, and the industrial internet. Their core objectives are to improve efficiency, enhance functionality, and achieve automated operation.

[0084] Oriented / Rotated Targets: Targets distributed at arbitrary angles in an image, whose compact bounding box is a rotated rectangle or any quadrilateral.

[0085] Rotating frame / rotating rectangle: A tilted rectangle represented by its center point, width, height, and angle or the coordinates of its four vertices.

[0086] One-hot hard labels: One-hot hard labels are a label representation method commonly used in machine learning classification tasks. A sample's label is typically represented as a sparse vector, with only one position marked as 1 and the rest as 0. This representation clearly indicates the category to which the sample belongs and is suitable for explicit single-label classification tasks.

[0087] Label Assignment: Label assignment is a technique for assigning target category labels to samples during the training phase. Optimizing label assignment strategies can significantly improve the training efficiency and performance of the model.

[0088] Semantic perception probability / semantic confidence: The soft probability that a sample point belongs to the foreground (or a certain category) in the image region, reflecting the strength of its "semantic representativeness" of the target.

[0089] NMS (Non-Maximum Suppression): A commonly used post-processing technique for deduplication, it filters out overlapping boxes sequentially based on candidate box scores and IoU.

[0090] Rotated IoU: Rotated IoU is an extension of traditional IoU, used to handle tasks involving the detection of rotated rectangles or objects. It is defined as the ratio of the intersection area to the union area of ​​two rotated rectangles.

[0091] Boundary discontinuity: The parameters of the rotated box are affected by the periodicity of the angle and the order of the vertices, resulting in parameter abrupt changes at the periodic boundaries or rule switching points, which leads to training instability and performance degradation.

[0092] The performance of classification and grading largely depends on the method of separating positive and negative samples. Positive and negative sample separation, also known as foreground / background point separation or label assignment, refers to the process of determining the foreground and background of the image regions corresponding to samples (anchor boxes or keypoints) on the output feature map during network training. This defines the foreground and background attributes of the sample, which are then used as positive and negative samples to train the model. The method of positive and negative sample separation defines the learning objective of the network model, directly determines the learning effect of the network model, and affects the final classification performance. Considering the characteristics of industrial images—large shape differences, blurred appearances, and complex backgrounds—more targeted positive and negative sample separation methods are of significant research importance for improving the performance of target detection in industrial images.

[0093] Existing technical solutions for separating positive and negative samples in target classification and grading training

[0094] (1) Hard labels based on human prior knowledge (anchor boxes / key points, see Figure 1 (b) Figure 1 In this context, 'a' represents the source image and the bounding box label. Most existing foreground / background point segmentation methods require direct definition based on manual prior knowledge, considering factors such as the target's location, shape, and scale. For example, anchor-box-based detection methods require manually setting anchor boxes of different shapes and scales, calculating their intersection-over-union (IoU) with the target boxes, and then using a threshold to filter and define the foreground / background. Keypoint-based detection methods directly define positive samples based on the target's location and scale. Meanwhile, improved methods such as dynamic segmentation (ATSS / OTA / PAA, etc.) have been proposed, requiring only relatively coarse manual prior knowledge to collect sufficient positive samples for all targets.

[0095] (2) Reweight the positive samples with probability / weight (GFL / VFL / FreeAnchor / AutoAssign methods, see...) Figure 1 (c) Based on the probability distribution of positive sample outputs, the attributes of positive samples are indirectly fine-tuned during training by adjusting the training weights of different positive samples, thereby completing the differentiated representation of the foreground target, avoiding the problem of treating samples equally, and improving detection accuracy.

[0096] The shortcomings of positive and negative sample separation in existing target classification and grading training

[0097] (1) Hard label assignment method: The positive and negative samples obtained by this method are not accurate enough in representing the foreground and background. On the one hand, the manual prior considers the regular shape of the target more. The defined positive samples are not accurate in representing the foreground of some tilted targets with a large aspect ratio and targets with special shapes (such as curved cables). They cannot be applied to all targets, resulting in the model's insufficient generalization ability to target shapes. Figure 1 In the illustration of b, most positive samples fall in areas unrelated to bananas. On the other hand, the image regions corresponding to the sample points mostly contain both foreground and background, with different foreground and background ratios. Therefore, these sample points have different representation abilities for the foreground and background. The one-hot annotation method ignores the differences in representation ability, treats all positive and negative samples equally, and uses them indiscriminately for model training, which also limits the classification accuracy of the model.

[0098] (2) Although the positive samples were reweighted with probability / weight to solve the problem of treating all positive and negative samples equally, the division of positive and negative samples is still based on artificial prior and still cannot solve the problem of generalization of the target shape.

[0099] The most commonly used post-processing method is Non-Maximum Suppression (NMS). Figure 2(As shown). NMS requires sequential filtering based on the detection box scores from largest to smallest. Due to the characteristics of industrial images—large directional changes and dense distribution of targets—the conventional NMS method is extremely inefficient in industrial scenarios. On the one hand, the number of targets in densely distributed scenes is extremely high, and the sequential processing adds a significant amount of additional time overhead; on the other hand, the calculation of the rotated IoU in directed target detection also leads to substantial time consumption. Therefore, the conventional NMS post-processing method is difficult to achieve high efficiency in directed target detection and has become one of the most time-consuming modules in the detection process, significantly reducing the overall detection speed.

[0100] Existing NMS acceleration and alternative technologies

[0101] (1) Parallelization / Matrix Acceleration (CUDA NMS / Fast NMS / Cluster NMS / Matrix NMS). This is the most commonly used acceleration strategy on parallel computing architecture hardware. It can directly achieve parallelization of IoU calculation, thereby effectively improving NMS efficiency.

[0102] (2) Approximation / Learning Replacement (Max Pooling Approximation, Learning-based NMS). One way to replace NMS is to use simple operators that can be processed in parallel, such as max pooling, to approximate the NMS processing; at the same time, there are also methods to design new network models to directly learn the NMS processing.

[0103] (3) One-to-one labeling strategies (OneNet / DeFCN / PSS, etc.) to weaken / eliminate NMS. By improving the training strategy, the model output can directly avoid duplicate detection results.

[0104] Disadvantages of existing NMS acceleration and alternatives

[0105] (1) Parallelization / Matrix Acceleration. Parallel computing-based acceleration methods are strictly dependent on hardware platforms and are not friendly to small hardware. At the same time, acceleration methods for serial filtering will suppress more boxes, increasing the probability of missed detection. Most importantly, parallel IoU calculation is only applicable to horizontal boxes, and it is difficult to effectively achieve parallel acceleration for rotated boxes, thus failing to improve the speed of rotated box detection.

[0106] (2) Approximation / learning substitution. These methods can only play an approximate role. Although they are faster, they also suffer from significant performance loss and cannot be effectively implemented in industrial scenarios.

[0107] (3) One-to-one labeling strategy. This strategy exacerbates the imbalance between positive and negative samples, making it difficult for the model to learn and achieve the same effect as NMS. Therefore, it is necessary to introduce an additional module into the model as an aid and guide. However, the final performance is still slightly lower than that of NMS.

[0108] In summary, existing acceleration methods either rely heavily on hardware platforms for parallel processing or require the introduction of additional network structure modules, and the detection performance generally decreases after acceleration.

[0109] Self-Learning Semantic Awareness (SemSL): This component constructs an "output confidence score" based on the classification and regression accuracy of the network output. Combined with initial values ​​from a multidimensional Gaussian engineering prior based on geometric and hierarchical consistency, it iteratively updates the semantic confidence score. Finally, it generates refined soft labels based on the semantic confidence score, achieving a transfer of labels from engineering priors to data adaptation. The technical principle of this component is as follows: Figure 3 As shown.

[0110] Progressive Optimal Single Sample (OSG): Within the regular positive sample set, a quality score is formed by fusing "predicted score × Gaussian prior," and the optimal single sample is selected. Then, progressive adjustment is made based on the degree of uniqueness enhancement of the quality prediction layer output (maximum-second largest difference) to ensure the optimal sample converges to be consistent with the high-score output. During inference, three-dimensional (spatial × hierarchical) max pooling nearest neighbor Gaussian soft suppression is applied to the quality × category output to enhance uniqueness and eliminate NMS. The technical principle of this component is as follows: Figure 4 As shown.

[0111] This application describes an efficient target extraction and high-precision classification and grading system for industrial intelligent connected scenarios, comprising components A, B, and C, specifically:

[0112] Component A: Self-learning semantic awareness-based label assignment (SemSL);

[0113] Based on the classification accuracy and localization accuracy of sample points obtained from the model output, an output confidence score is constructed for each sample. This score gradually approximates the corresponding semantic perception probability during model learning, guiding the generation of foreground and background attribute labels and achieving self-learning. At the beginning of training, since there is no reliable semantic perception probability distribution, a manually defined multidimensional Gaussian distribution is used as the initial value for the semantic perception probability. During training, this initial value is updated based on the constructed output confidence score, and foreground and background attribute labels are generated based on the updated semantic perception probability. This process continues iteratively, and the foreground and background attribute labels gradually become entirely determined by the output confidence score, causing the constructed output confidence score to gradually approach the semantic perception probability, thus achieving self-learning. Finally, the confidence score constructed from the output accuracy effectively represents the semantic perception probability of the sample point, and the foreground and background attribute labels generated under this confidence score have a more accurate representation of the target. The specific implementation steps are as follows:

[0114] (1) Construction of output confidence

[0115] Classification accuracy Directly take the network output class confidence score; if it includes a quality branch, multiply the two scores; regression accuracy. Based on prediction boxes With target box of To enhance the discrimination of high-quality samples, normalization is performed to increase variance. The formula is as follows:

[0116] ;

[0117] If the objective is optimal If the value is ≥0.5, then the optimal sample regression accuracy is approximately 1, and the accuracy of other samples follows the trend. Reduce rapid decline, based on classification accuracy and regression accuracy The output confidence score is obtained using the following formula:

[0118] ;

[0119] The output confidence score can gradually approach the semantic perception probability during training, where γ is used to adjust the relative weight ratio of the two branches.

[0120] (2) Multidimensional Gaussian initial value, i.e. semantic prior initialization

[0121] Initial values ​​are generated only for samples falling into the target region; in rotated scenarios, the interior is determined by the rotation coordinates.

[0122] ;

[0123] in, and From sample to target The center offset; This is the step size for that layer; This represents the distance from the sample to each side of the target. Control level soft matching; , These represent the spatial and hierarchical variances, respectively, and the samples that fall outside the target. See the detailed diagram. Figure 5 The "a" is a punctuation mark.

[0124] (3) Semantic probability update, i.e., self-learning

[0125] Will and Perform adaptive fusion to obtain semantic confidence. Real-time approximation of semantic probability distribution:

[0126] ;

[0127] in, The system adaptively adjusts based on the highest output confidence level under the current objective, shifting from initial prior dominance to later model dominance. This is for fine-tuning the index.

[0128] (4) Generation of fine-grained soft tags

[0129] For each target g, the sample set Normalization and nonlinear compression yield the final soft tags. (In the case of a single class):

[0130] ;

[0131] in, The steepness of the control curve is controlled; the normalization process ensures that the maximum label for each target is 1, preserving the continuity and diversity of soft labels. The specific function curve is shown below. Figure 5 As shown in b.

[0132] (5) Training and Loss

[0133] Classification loss:

[0134] ;

[0135] Regression loss, GIoU or IoU loss, regression only weights samples where y > 0.

[0136] ;

[0137] Quality Branch: Labels are acceptable The loss is calculated using binary cross-entropy:

[0138] ;

[0139] Total loss:

[0140] ;

[0141] Where N is the number of positive weighted samples.

[0142] Component B: Optimal Single-Sample Asymptotic Assignment and NMS Elimination (OSG)

[0143] Based on the target location and model prediction accuracy, the optimal positive sample is further selected from multiple positive samples obtained by conventional label allocation methods and used as the label for the quality prediction layer. As training progresses, the high-scoring sample points output by the quality prediction layer gradually satisfy uniqueness. To alleviate the inconsistency between these high-scoring sample points and the assigned optimal positive sample, a progressive allocation strategy based on the output of the quality prediction layer is introduced in the optimal single-sample label allocation process. This ensures that the spatial positions of the assigned optimal positive sample and the high-scoring sample points output by the quality prediction layer gradually align during training. Finally, a three-dimensional cube suppression strategy is proposed during the model inference stage to quickly suppress local non-maximums, further strengthening the one-to-one correspondence between detection results and targets. Introducing the optimal single-sample matching strategy-based progressive label allocation (OSG) scheme during training effectively improves the one-to-one correspondence between output detection results and targets, effectively avoids duplicate detections, completely eliminates the NMS post-processing process, and improves detection speed. The specific implementation steps are as follows:

[0144] (1) Positive sample quality score (initial selection of the best)

[0145] The set of positive samples obtained from regular label assignment In the middle, the quality score is constructed by fusing the predicted score with the Gaussian prior. :

[0146] ;

[0147] in The generation method is the same as component A. Initial value from Dominated, later by leading, To measure the overall representativeness of the sample to the target;

[0148] Initial selection of the optimal sample: .

[0149] (2) Gradual adjustment (convergence consistent with quality output)

[0150] Let the quality branch output be... If the difference between the maximum and second-largest values ​​increases, it indicates an improvement in uniqueness, and the value should be... Gradually "pulled" to the maximum Location:

[0151] ;

[0152] in To adjust the intensity; The bigger it is, the more you trust it. Guidance; only in the mid-to-late stages When the value is large, this factor should be increased to avoid initial instability; a visual diagram of the gradual adjustment process is shown below. Figure 6 As shown.

[0153] (3) Quality layer one-to-one label and loss

[0154] For each target Only set As the only positive sample, its label The labels for all other positive and negative samples are 0;

[0155] The loss corresponding to the quality layer (cross-entropy or BCE): ;

[0156] The classification and regression layers still use the conventional (multiple positive sample labels) allocation to avoid extreme imbalance.

[0157] (4) Lightweight suppression replacing NMS during inference (three-dimensional cube)

[0158] A schematic diagram of the three-dimensional maximum pooling cube suppression process is shown below. Figure 7 As shown. The specific implementation process is as follows:

[0159] Constructing the suppressed input score: n≈0.4 (improving focus on high scores);

[0160] 3D max pooling (spatial × hierarchical FPN), kernel = 3×3×3, stride = 1, boundary padding = 1; upsample each layer to the same scale (repeated by pixels), then fall back;

[0161] Mask generation hard suppression: Otherwise, it is 0;

[0162] Mask generation Gaussian soft suppression: ;

[0163] Output score: Select Top-K as the final result, and do not perform NMS again;

[0164] Component C: Rotated Rectangle Regression Based on GWD Loss

[0165] To improve the positioning accuracy of the rotated bounding box, the existing Gaussian Wasserstein DistanceRegression Loss (GWD Loss) is used to regress the rotated rectangle. First, a two-dimensional Gaussian distribution is used to model the rotated target (see...). Figure 8 The specific formula is as follows:

[0166] ;

[0167] Therefore, the Gaussian Wasserstein Distance (GWD) is used to replace the non-differentiable rotational IoU, as shown in the following formula:

[0168] ;

[0169] The Loss value is calculated based on GWD using the following formula:

[0170] ;

[0171] This solution addresses the boundary discontinuity problem in the training process of rotated bounding box regression at its root, without introducing any modules or increasing the number of parameters in the rotated bounding box representation, thereby improving positioning accuracy while maintaining low parameters.

[0172] Finally, components A, B, and C were integrated and industrialized.

[0173] Classification + Localization (SemSL + GWD): Significantly improves AP in rotating scenes, especially AP75 at high IoU thresholds; SemSL improves classification stability in complex scenes, and GWD improves regression accuracy of rotated bounding boxes.

[0174] End-to-end (SemSL+OSG): OSG is introduced during training, and 3D cube suppression is used to completely remove NMS during inference; significantly improving frame rate on industrial edge devices. The algorithm optimized for OSG based on SemSL is simpler than the original OSG method. The original OSG method requires constructing positive sample quality scores based on target location and model output accuracy to complete the optimal single-sample assignment process. After the improvement based on SemSL, the positive sample quality scores can be directly replaced by the directed target semantic confidence generated by SemSL. Furthermore, it can be directly combined with the progressive assignment strategy in OSG to complete a self-learning-based optimal single-sample progressive label assignment method, achieving an effective combination of SemSL and OSG.

[0175] High-precision, high-efficiency, and lightweight system (SemSL+GWD+OSG): The three components can be seamlessly integrated into any target detection method. This invention migrates the above modules to the RTMDet architecture. The core is to adopt a lightweight backbone and PAFPN pyramid module, combined with MixUp / Mosaic, optimizer and learning rate scheduling, to obtain a high-precision, high-efficiency, and lightweight detection system for industrial intelligent networking. It can be deployed in small embedded hardware to realize edge-side and end-side applications.

[0176] Component A: An Alternative to SemSL

[0177] One type of method that uses purely dynamic positive and negative samples (such as ATSS / OTA / PAA methods) has limited performance improvement.

[0178] Another type of method only uses IoU or confidence weighting (such as GFL / VFL / FreeAnchor methods) and does not explicitly approximate the true semantic probability.

[0179] The SemSL technology proposed in this invention integrates semantic self-learning and Gaussian soft starting, resulting in greater stability and better generalization.

[0180] Component B: An alternative to OSG

[0181] Performing a one-to-one classification directly at the classification layer will trigger extreme imbalance. Therefore, related methods need to introduce additional structures (DeFCN, PSS, GCN), which increases the computational cost and limits the flexibility of algorithm deployment.

[0182] This invention performs a one-to-one comparison only at the quality layer, while classification / regression follows the conventional allocation, avoiding structural complexity and making training more stable. At the same time, the proposed 3D cube suppression in the inference stage has lower complexity than NMS and is easier to parallelize, especially in rotated box detection, where it has a more obvious advantage.

[0183] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A high-efficiency target extraction and high-precision classification and grading system for industrial intelligent connected scenarios, characterized in that, Includes component A, component B, and component C, specifically: Component A: Label allocation based on self-learning semantic awareness. The specific implementation steps are as follows: (1) Construction of output confidence Classification accuracy Directly take the network output class confidence score; if it includes a quality branch, multiply the two scores; regression accuracy. Based on prediction boxes With target box of To enhance the discrimination of high-quality samples, normalization is performed to increase variance. The formula is as follows: ； If the objective is optimal If the value is ≥0.5, then the optimal sample regression accuracy is approximately 1, and the accuracy of the remaining samples follows the trend. Reduce rapid decline, based on classification accuracy and regression accuracy The output confidence score is obtained using the following formula: ； The output confidence score can gradually approximate the semantic perception probability during training, where γ is used to adjust the relative weight ratio of the two branches; (2) Multidimensional Gaussian initial value, i.e. semantic prior initialization Initial values ​​are generated only for samples falling into the target region; in rotated scenarios, the interior is determined by the rotation coordinates. ； in, and From sample to target The center offset; This is the step size for that layer; This represents the distance from the sample to each side of the target. Control level soft matching; , These represent the spatial and hierarchical variances, respectively, and the samples that fall outside the target. ; (3) Semantic probability update, i.e., self-learning Will and Perform adaptive fusion to obtain semantic confidence. Real-time approximation of semantic probability distribution: ； in, The system adaptively adjusts based on the highest output confidence level under the current objective, shifting from initial prior dominance to later model dominance. To fine-tune the index; (4) Generation of fine-grained soft tags For each target g, the sample set Normalization and nonlinear compression yield the final soft tags. In the case of a single class: ； in, The steepness of the control curve is controlled; the normalization process ensures that the maximum label of each target is 1, preserving the continuity and difference of soft labels; (5) Training and Loss Classification loss: ； Regression loss, GIoU or IoU loss, regression only weights samples where y > 0. ； Quality Branch: Labels are acceptable The loss is calculated using binary cross-entropy: ； Total loss: ； Where N is the number of positive weighted samples; Component B: Optimal single-sample progressive allocation and NMS elimination, the specific implementation steps are as follows: (1) Positive sample quality score, initial selection of the best The set of positive samples obtained from regular label assignment In the middle, the quality score is constructed by fusing the predicted score with the Gaussian prior. : ； in The generation method is the same as component A. Initial value from Dominated, later by leading, To measure the overall representativeness of the sample to the target; Initial selection of the optimal sample: ; (2) Gradual adjustment, consistent with quality output convergence Let the quality branch output be... If the difference between the maximum and second-largest values ​​increases, it indicates an improvement in uniqueness, and the value should be... Gradually "pulled" to the maximum Location: ； in To adjust the intensity; The bigger it is, the more you trust it. Guidance; only in the mid-to-late stages When the value is large, strengthen this item to avoid initial instability; (3) Quality layer one-to-one label and loss For each target Only set As the only positive sample, its label The labels for all other positive and negative samples are 0; The loss corresponding to the quality layer is cross-entropy or BCE: ; The classification and regression layers still use the conventional multiple positive sample label allocation to avoid extreme imbalance. (4) Lightweight suppression replacing NMS during inference period, three-dimensional cube The three-dimensional max-pooling cube suppression process is implemented as follows: Constructing the suppressed input score: n≈0.4 (improving focus on high scores); A 3D max pooling space × layer FPN is used, with kernel = 3×3×3, stride = 1, and boundary padding = 1. Each layer is upsampled to the same scale and then backsampled. Mask generation hard suppression: Otherwise, it is 0; Mask generation Gaussian soft suppression: ; Output score: Select Top-K as the final result, and do not perform NMS again; Component C: Rotated Rectangle Regression Based on GWD Loss To improve the positioning accuracy of the rotated bounding box, the existing Gaussian Wasserstein Distance Regression Loss is used to regress the rotated rectangle. First, a two-dimensional Gaussian distribution is used to model the rotated target, and the specific formula is as follows: ； Therefore, the Gaussian Wasserstein Distance (GWD) is used to replace the non-differentiable rotational IoU, as shown in the following formula: ； The Loss value is calculated based on GWD using the following formula: ； Finally, components A, B, and C were integrated and industrialized.

2. The efficient target extraction and high-precision classification and grading system for industrial intelligent connected scenarios as described in claim 1, characterized in that, Component A uses purely dynamic positive and negative samples.

3. The efficient target extraction and high-precision classification and grading system for industrial intelligent connected scenarios as described in claim 1, characterized in that, Component A only uses IoU or confidence weighting to explicitly approximate the true semantic probability.

Images