Video data normativity review method and system based on artificial intelligence

Abstract

The invention relates to the technical field of video detection, in particular to a video data normativity review method and system based on artificial intelligence, and the method comprises the steps: collecting a field image / video, and marking the field image / video according to a standard category; the method comprises the following steps: constructing an SC-YOLO model containing a CSPDarnet (C3k2, SPPF, C2PSA), an FPN-PAN neck part and a depth separable multi-scale detection head, and constructing the SC-YOLO model containing the CSPDarnet (C3k2, SPPF, C2PSA). Constructing a target function weighted by frame regression, classification and distributed focus loss by forward output and annotation, solving a gradient and implementing coordinate-level preconditions, amplitude clipping and weight attenuation by matching Hutchinson or GNB diagonal curvature estimation with EMA smoothing, and circularly updating to convergence; and deploying the optimization model at an acquisition end / edge device, and outputting specification categories, positions and confidence information in real time for security management. According to the invention, the key problems of easy leak detection, unstable training convergence, insufficient efficiency, limited real-time deployment at the edge side and the like of small target / shielding / transparent materials in the prior art can be solved, so that the manual inspection error and time delay are reduced.

Description

Technical Field

[0001] This invention relates to the field of video inspection technology, and in particular to a method and system for reviewing the standardization of video data based on artificial intelligence. Background Technology

[0002] In high-risk scenarios such as engineering construction, on-site safety relies on continuous monitoring of whether personnel are wearing personal protective equipment (such as safety helmets, reflective vests, gloves, and goggles) according to regulations. However, traditional manual inspections suffer from high labor intensity, high error rates, and poor coverage continuity, making it difficult to meet the real-time compliance verification needs of large-scale and complex working conditions. One-stage target detection, represented by YOLO, has achieved a good balance between real-time performance and accuracy, but it still faces challenges such as missed detections and unstable positioning for small, obscured, or transparent PPE components. At the same time, cluttered backgrounds and scale diversity pose challenges to the stability and efficiency of training convergence. Summary of the Invention

[0003] In view of the above technical problems, the present invention provides a video data standardization review method and system based on artificial intelligence to solve key problems in the prior art, such as easy omission of small targets / occlusion / transparent materials, unstable training convergence and insufficient efficiency, and limited real-time deployment at the edge, thereby reducing manual inspection errors and time delays.

[0004] Other features and advantages of this disclosure will become apparent from the following detailed description, or may be learned in part from practice of this disclosure.

[0005] According to one aspect of the present invention, a video data normative review method based on artificial intelligence is proposed, the method comprising: Collect images or video frames from the shooting scene, label the targets according to the standardized categories, and obtain training samples; A detection model based on SC-YOLO is constructed. The detection model includes: feature extraction using the CSPDarknet backbone network, in which C3k2 and SPPF modules are set, and a C2PSA attention module is set at the tail of the CSPDarknet backbone network; upsampling, connecting, and fusing multi-scale features using a neck structure of feature pyramid and path aggregation; and outputting class probability, bounding box position distribution, and box quality index using a multi-scale detection head. The multi-scale detection head uses depthwise separable convolution to reduce computational cost. The labeled training samples are input into the detection model, and multi-level semantic and texture features are extracted by the CSPDarknet backbone network. After multi-scale semantic fusion is performed by the neck structure, the samples are input into the multi-scale detection head, and the class probability and bounding box distribution of each candidate position are output. Based on the category probabilities, bounding box distribution, and annotation results of the forward output, a weighted comprehensive loss is constructed, which includes bounding box regression loss, classification loss, and distributed focus loss, to simultaneously constrain positioning accuracy, category determination, and bounding box quality. The gradient vector of the weighted comprehensive loss is obtained for all trainable parameters of the detection model, and the local curvature of the objective function formed by the weighted comprehensive loss is estimated. The curvature estimation includes: obtaining the diagonal curvature vector using a diagonal Hessian estimator formed by the inner product of a random vector and the gradient or a Gauss-Newton-Bartlett estimator based on the model's log odds output and sampling labels, and performing exponential moving averages on the gradient vector and the curvature vector in the time dimension to suppress noise. The gradient vector is scaled by the curvature vector to form a preconditional gradient according to the coordinates. The coordinate magnitude is clipped on the preconditional gradient. When the curvature vector is estimated to be zero, an update based on momentum sign is adopted, and weight decay is combined to suppress overfitting. Based on this, all trainable parameters of the detection model are updated. The process of inputting the training samples into the detection model and updating all trainable parameters based on the precondition gradient is repeated until the convergence condition is met, thus obtaining the optimized detection model. The optimized detection model is deployed on an image acquisition terminal or edge computing device to perform real-time inference on the input from the shooting scene and output standardized detection results containing the standardized category, location and confidence information for security management.

[0006] Furthermore, the CSPDarknet backbone network employs cross-stage partial connections to divide the feature mapping into a preservation path and a transformation path, which are then merged in subsequent stages to enhance gradient flow and reduce redundant computation. The C3k2 module is used to improve feature representation efficiency while maintaining the receptive field. The SPPF module is used to aggregate multi-scale contextual information. The C2PSA attention module is used to suppress background interference and highlight salient regions related to canonical categories.

[0007] Furthermore, the neck structure, based on the feature pyramid and path aggregation, is executed sequentially as follows: High-level semantic features are upsampled, and the upsampled results are connected and fused with low-level detail features of adjacent levels. The C3k2 module is then used for nonlinear recombination to form multi-scale fused features that include both semantics and details.

[0008] Furthermore, the multi-scale detection head feeds the multi-scale fused features output from the neck structure into parallel prediction branches, which employ depthwise separable convolutions to reduce computational load and output the following respectively: The probability distribution for class determination, the center and boundary distribution for bounding box localization, and the score for box quality measurement are used to adapt to the standard categories of different sizes.

[0009] Furthermore, the weighted comprehensive loss is calculated by: first calculating the bounding box regression loss to penalize the positional difference between the predicted box and the labeled box; then calculating the classification loss to measure the consistency between the class probability and the true class; simultaneously calculating the distributed focus loss to improve the fineness of the boundary distribution; and finally linearly combining the bounding box regression loss, the classification loss, and the distributed focus loss with preset weights to obtain the objective function used for backpropagation.

[0010] Furthermore, when implementing the coordinate magnitude clipping, the following steps are included: The magnitude of each coordinate update after curvature scaling is compared with the threshold, and the coordinate update exceeding the threshold is compressed to the threshold range. If the curvature vector of a certain coordinate is estimated to be zero or negative, the sign update based on the historical momentum direction is used in the corresponding coordinate to maintain training stability, and the weight decay is applied to the parameters before each parameter update to limit the unbounded growth of the parameter magnitude.

[0011] According to a second aspect of this disclosure, an artificial intelligence-based video data normative review system is provided, the system comprising: The acquisition module is used to acquire images or video frames from the shooting scene, label the targets according to the standard categories, and obtain training samples; The model building module is used to construct a detection model based on SC-YOLO. The detection model includes: feature extraction using the CSPDarknet backbone network, which includes a C3k2 module and an SPPF module, and a C2PSA attention module at the tail of the CSPDarknet backbone network; upsampling, connecting, and fusing multi-scale features using a neck structure of feature pyramids and path aggregation; and outputting class probabilities, bounding box position distributions, and box quality indices using a multi-scale detection head, wherein depthwise separable convolutions are used in the multi-scale detection head to reduce computational cost. The feature extraction module is used to input the labeled training samples into the detection model, extract multi-level semantic and texture features through the CSPDarknet backbone network, perform multi-scale semantic fusion through the neck structure, and input the samples into the multi-scale detection head, outputting the class probability and bounding box distribution of each candidate position. The comprehensive loss construction module is used to construct a weighted comprehensive loss including bounding box regression loss, classification loss and distributed focus loss based on the category probability, bounding box distribution and annotation results of the forward output, so as to simultaneously constrain the positioning accuracy, category determination and bounding box quality. The second-order estimation module is used to obtain the gradient vector of the weighted comprehensive loss for all trainable parameters of the detection model, and to estimate the local curvature of the objective function formed by the weighted comprehensive loss. The curvature estimation includes: obtaining the diagonal curvature vector using a diagonal Hessian estimator formed by the inner product of a random vector and the gradient or a Gauss-Newton-Bartlett estimator based on the model's log odds output and sampling labels, and performing exponential moving averages in the time dimension on the gradient vector and the curvature vector respectively to suppress noise. The parameter update module is used to scale the gradient vector by the curvature vector to form a preconditional gradient, perform coordinate magnitude clipping on the preconditional gradient to limit anomalous steps, and degenerate into a momentum sign-based update when the curvature vector estimate is zero or unreliable, and combine it with weight decay to suppress overfitting, thereby updating all trainable parameters of the detection model. The iterative training module is used to repeatedly execute the process of inputting the training samples into the detection model and updating all trainable parameters based on the precondition gradient until the convergence condition is met, thereby obtaining the optimized detection model. The deployment module is used to deploy the optimized detection model on the image acquisition terminal or edge computing device, perform real-time inference on the input from the shooting scene, and output standardized detection results containing the standardized category, location and confidence information for security management.

[0012] The technical solution disclosed herein has the following beneficial effects: It can achieve more robust detection. Multi-scale feature fusion and C2PSA attention help maintain the identification and location of small-sized and transparent PPE in crowded, occluded and complex backgrounds. The recall capability is more suitable for safety and compliance scenarios. Training is more reliable. Sophia’s Hessian diagonal estimation and element-level pruning bring curvature-adaptive updates and degradation protection, which alleviate gradient oscillations and accelerate stable convergence. More efficient deployment: The detection head uses depthwise separable convolution combined with a lightweight backbone, balancing inference speed and accuracy, making it suitable for real-time review at resource-constrained edge devices. It is more generalizable and practical, maintaining a consistent performance trend across different datasets and diverse field conditions, supporting scenario migration from fixed factory areas to dynamic construction sites, and thus providing continuous, quantitative, and traceable normative reviews for safety management. Attached Figure Description

[0013] Figure 1 This is a flowchart of an artificial intelligence-based video data compliance review method as described in the embodiments of this specification; Figure 2This is a structural block diagram of an artificial intelligence-based video data standardization review system as described in the embodiments of this specification. Detailed Implementation

[0014] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided to make this disclosure more comprehensive and complete, and to fully convey the concept of the example embodiments to those skilled in the art. The described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. In the following description, numerous specific details are provided to give a full understanding of embodiments of this disclosure. However, those skilled in the art will recognize that the technical solutions of this disclosure can be practiced with one or more of the specific details omitted, or other methods, components, apparatus, steps, etc., can be employed. In other instances, well-known technical solutions are not shown or described in detail to avoid obscuring various aspects of this disclosure.

[0015] Furthermore, the accompanying drawings are merely illustrative of this disclosure. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0016] This invention provides a method for reviewing the compliance of video data for products based on artificial intelligence. (Refer to...) Figure 1 The diagram shown is a flowchart illustrating a video data compliance review method based on artificial intelligence, according to an embodiment of the present invention. This method can be applied to electronic devices such as personal computers, servers, controllers, and host computers. The method can be executed by a device, which can be implemented by software and / or hardware. Specifically, the method may include the following steps S101-S108: In step S101, images or video frames of the shooting scene are acquired, and the targets are labeled according to the standard categories to obtain training samples.

[0017] Before labeling, the scope and definition of the standard category must be determined. This example demonstrates a construction safety compliance scenario, with the standard category including at least the categories of personal protective equipment (PPE) worn / worn and the objects most closely interacting with it. These categories are used to train the detection model to identify elements related to compliance status. Specifically, a category system based on two complementary data sources can be used: one includes six categories such as boots, safety glasses, gloves, helmets, people, and vests; the other includes three categories such as safety helmets, reflective vests / work clothes, and workers. These categories cover transparent materials (such as safety glasses) and small, deformable components (such as gloves), and include personnel / workers in the contextual target to support the subsequent modeling requirements of scenario semantics and the human-equipment relationship for regulatory review tasks.

[0018] The data acquisition side primarily utilizes camera infrastructure, including fixed or mobile camera deployments. Continuous monitoring acquires on-site images or video streams, ensuring coverage of different processes, varying crowd densities, and time periods, thus creating diverse training sample sources. To guarantee sample representativeness and generalization, the acquired content should ideally cover indoor / outdoor environments, various lighting conditions, and complex backgrounds. In practical implementation, video sources with continuous monitoring capabilities can provide material for extracting keyframes from the video stream. Simultaneously, the camera infrastructure and computing resources constitute the deployment foundation and data entry point for the automated detection system.

[0019] Specifically, during annotation, each image or extracted video frame is checked one by one to see if the aforementioned standardized category targets exist. Once a target is found, a bounding box is used to delineate the instance on the image, and it is labeled with a standardized category label. To ensure the computability of localization and quality assessment in subsequent training, the annotation must meet the structured requirement of pairing categories and bounding boxes; the annotation should highlight the accurate location and clear category attribution of the bounding box, so that the model can simultaneously learn the three aspects of location, category, and box quality under supervised signals.

[0020] The training samples obtained through the above collection and annotation process not only reflect the environmental diversity of the construction site, but also meet the category and bounding box standards required for target detection training in terms of structured annotation. Through integrated control of class distribution, scene conditions and data partitioning, the sample set can provide reliable input for subsequent model training and evaluation, and support the learning needs of the normative review task for small targets, occlusion and transparent boundaries in complex scenes containing elements such as personnel, reflective vests, safety helmets, gloves, safety glasses and boots.

[0021] In step S102, a detection model based on SC-YOLO is constructed. The detection model includes: feature extraction using the CSPDarknet backbone network, in which C3k2 and SPPF modules are set, and a C2PSA attention module is set at the tail of the CSPDarknet backbone network; upsampling, connecting, and fusing multi-scale features using a neck structure of feature pyramids and path aggregation; and outputting class probabilities, bounding box position distributions, and box quality indicators using a multi-scale detection head, in which depthwise separable convolution is used to reduce computational cost.

[0022] The CSPDarknet backbone network employs cross-stage partial connections to divide feature mapping into a preservation path and a transformation path, which are then merged in subsequent stages to enhance gradient flow and reduce redundant computation. The C3k2 module is used to improve feature representation efficiency while maintaining the receptive field. The SPPF module is used to aggregate multi-scale contextual information. The C2PSA attention module is used to suppress background interference and highlight salient regions related to canonical categories.

[0023] Furthermore, the neck structure based on the feature pyramid and path aggregation sequentially performs the following: upsampling of high-level semantic features, connecting and fusing the upsampling results with low-level detail features of adjacent levels, and performing nonlinear recombination through the C3k2 module to form multi-scale fused features containing semantics and details.

[0024] In step S102, the detection model is first constructed as an integrated architecture consisting of a backbone network, a neck structure, and a multi-scale detection head. To meet the efficiency-accuracy trade-off between real-time review and edge deployment, a lightweight framework based on YOLOv11n is selected as the backbone paradigm. This paradigm uses a robust feature extraction backbone, a neck structure that performs multi-scale fusion around the feature pyramid, and a detection head that is simultaneously oriented towards classification and localization as three major components to work together, thereby balancing speed and accuracy in complex scenarios.

[0025] The backbone network adopts CSPDarknet, integrating two key modules, C3k2 and SPPF, into its hierarchical structure. A C2PSA attention mechanism is also introduced at the tail of the backbone. CSPDarknet strategically partitions and re-aggregates feature maps through cross-stage partial connections to reduce computational redundancy and optimize gradient propagation while preserving semantic information integrity; this makes it a suitable choice for real-time detection tasks under computationally limited conditions. C3k2 (Convolutional Cross-Connections with kernel size 2) is used to enhance representation capabilities at multiple scales, while SPPF (Spatial Pyramid Pooling Fast) is used to expand the effective receptive field and aggregate multi-scale context; both work together to improve the feature extraction quality of the backbone. C2PSA (Cross-Stage Partial Self-Attention) emphasizes salient regions and suppresses irrelevant information at the tail of the backbone, thereby enhancing context sensitivity in densely populated and occluded scenes, thus facilitating robust recognition of canonical category-related elements. The structure and mechanism of the aforementioned backbone ensure that semantically rich and hierarchically clear feature inputs can be continuously provided in the subsequent multi-scale fusion and prediction stages.

[0026] The neck structure revolves around a feature pyramid and path aggregation strategy, proceeding in the order of upsampling, connection, and recombination: first, high-level semantic features are upsampled; then, they are connected channel-wise with low-level detail features from adjacent layers to aggregate cross-layer information; finally, non-linear recombination and integration are achieved through C3k2, forming a multi-scale fusion feature that combines high-level semantics and low-level texture. Structurally, this process is guided by lightweight computation, ensuring efficient transmission of semantics and details at different resolutions and scales. The two ends provide stable multi-scale semantic bridges for the backbone output and the detection head input.

[0027] Additionally, SC-YOLO is a real-time detection variant based on YOLOv11n. It uses the CSPDarknet backbone and integrates the second-order optimizer Sophia for PPE compliance detection in resource-constrained environments, balancing accuracy and latency.

[0028] In step S103, the labeled training samples are input into the detection model, multi-level semantic and texture features are extracted by the CSPDarknet backbone network, and multi-scale semantic fusion is performed by the neck structure before being input into the multi-scale detection head, outputting the class probability and bounding box distribution of each candidate position.

[0029] The multi-scale detection head feeds the fused features into parallel prediction branches, which employ depthwise separable convolutions to reduce computation and output: a probability distribution for class determination, a center and boundary distribution for bounding box localization, and a score for box quality measurement, to adapt to canonical classes of different sizes.

[0030] In step S103, the labeled training samples are input into the detection model. In this step, only the image tensors in the samples are used for calculation. The model first extracts multi-level semantic and texture features in the backbone, then performs multi-scale semantic fusion in the neck, and finally generates output tensors for subsequent supervised calculations at the multi-scale detection head.

[0031] The multi-scale detection head uses parallel resolution branches to perform prediction tasks, outputting probabilities for class determination, regression values ​​or location distributions for bounding box localization, and predictions adapted to box quality assessment. To reduce computational overhead and improve throughput without sacrificing accuracy, depthwise separable convolutions are introduced into each branch of the detection head. This detection head sequentially interfaces with the aforementioned neck output, allowing high-resolution branches to prioritize fine-grained targets while low-resolution branches carry global semantics, thus achieving robust perception of standard class targets at multiple scales. Its output format corresponds one-to-one with the subsequent training objectives used for localization accuracy, class determination, and box quality assessment, ensuring structure-object consistency and trainability.

[0032] Through the organic combination of the aforementioned backbone, neck, and detection head, the constructed SC-YOLO detection model is structurally characterized by: maintaining efficient gradients and semantic fidelity through cross-stage partial connections; ensuring multi-scale context aggregation through C3k2 and SPPF; suppressing irrelevant backgrounds and emphasizing key regions through C2PSA; and simultaneously completing classification and localization outputs at multiple resolutions using a lightweight detection head. This overall design achieves a balance between efficiency and accuracy in complex environments.

[0033] The forward propagation in step S103 is limited to the computational path of input image, backbone feature extraction, neck multi-scale fusion, and parallel prediction by the detection head. The backbone provides a multi-level semantic foundation through CSPDarknet and its supporting modules. Sampling and connection above the neck achieve cross-layer alignment of semantics and details. The detection head produces class probabilities, bounding box position distributions, and box quality indicators on multi-resolution branches, laying the data foundation for subsequent steps to construct supervision signals based on these forward outputs and annotation results.

[0034] In step S104, based on the category probabilities, bounding box distribution and annotation results of the forward output, a weighted comprehensive loss including bounding box regression loss, classification loss and distributed focus loss is constructed to simultaneously constrain positioning accuracy, category determination and bounding box quality.

[0035] The weighted comprehensive loss is calculated by first calculating the bounding box regression loss to penalize the positional difference between the predicted box and the labeled box, then calculating the classification loss to measure the consistency between the class probability and the true class, and simultaneously calculating the distributed focus loss to improve the fineness of the boundary distribution. Finally, the bounding box regression loss, the classification loss and the distributed focus loss are linearly combined with preset weights to obtain the objective function used for backpropagation.

[0036] In step S104, based on the forward output and annotation results obtained in step S103, a weighted comprehensive loss is constructed as the training objective. This comprehensive loss consists of three parts: bounding box regression loss for positioning accuracy. Classification loss used for category determination and distributed focus loss used for frame quality assessment. The total loss is then summed using a linear weighted average. Its formal expression is: ; in, The focus is on ensuring the consistency of the positions of the predicted bounding box and the labeled bounding box. Used for determining the target category. The quality of the bounding box is measured by modeling and penalizing the boundary distribution; the aforementioned weighting coefficients are determined empirically through pre-experimentation to achieve an optimal balance between precision and recall, where the weighting coefficients are... Giving it higher weight to highlight its importance in regulatory review scenarios, Give moderate weights to emphasize boundary quality, while... Lower weights are assigned to reflect the relatively easy class distinction. This comprehensive loss serves as the objective function for subsequent backpropagation, providing a unified basis for gradient calculations across all layers of the network.

[0037] The specific calculation process follows the order of first dividing into items and then summarizing: first, the positional difference between the predicted boundary and the labeled boundary is used to construct... This is used to penalize positioning errors and push for geometric alignment of the bounding box; then, a consistency metric between the predicted class probability and the true class is obtained. This improves the accuracy of category determination; simultaneously, it constructs a boundary distribution based on the output of the detection head. By supervising the shape and kurtosis of the boundary distribution, the sensitivity to boundary quality and fine-grained localization is improved. After the three sub-losses are calculated and normalized, they are linearly combined according to preset weights to obtain... This loss decomposition and weighting framework is used to balance the contributions of localization, classification, and bounding box quality, ensuring stable detection performance and trainability under different scales and complex backgrounds.

[0038] In step S105, the gradient vector of the weighted comprehensive loss is obtained for all trainable parameters of the detection model, and the local curvature of the objective function formed by the weighted comprehensive loss is estimated. The curvature estimation includes obtaining the diagonal curvature vector using a diagonal Hessian estimator formed by the inner product of a random vector and the gradient or a Gauss-Newton-Bartlett estimator based on the model's log-odds output and sampling labels, and performing exponential moving averages over time on the gradient vector and the curvature vector respectively to suppress noise.

[0039] Specifically, in implementing step S105, the gradient vectors of all trainable parameters of the detection model are first calculated using the weighted comprehensive loss as the objective function. Let all trainable parameters be denoted as... The weighted comprehensive loss is The gradient at the t-th training step is calculated in the following form: ; in, This represents the partial derivative operator with respect to all parameters of the network. The set of trainable parameters covering the trunk, neck, and detection head layers is given by this formula. This serves as the foundation for subsequent curvature estimation and optimization steps.

[0040] In the curvature estimation section, to obtain the diagonal curvature vector for each coordinate precondition, this embodiment employs two efficient diagonal Hessian estimation approaches, one of which is a Hutchinson-type estimator. The specific process is as follows: For the current parameter Calculate a small batch loss ;from Draw a random vector u; return ,in This indicates element-wise multiplication. Let be the inner product. This estimate is unbiased and satisfies ... That is, its expectation is equal to the diagonal elements of the Hessian, thus obtaining a coordinate-wise curvature approximation without explicitly constructing an integer matrix. The chained operations described above, which calculate mini-batch loss, random vector sampling, gradient, inner product, and element-wise multiplication, can obtain the diagonal curvature vector synchronously during training with low additional overhead.

[0041] The second is the Gauss-Newton-Bartlett (GNB) estimator, suitable for classification tasks. It utilizes the cross-entropy structure and Bartlett's identity to efficiently approximate the diagonal of the Gauss-Newton matrix. The specific steps are as follows: From the current small batch input Calculate the corresponding logarithmic odds ;according to Generate a label for each sample; calculate ;return This construction, without explicitly forming a Hessian, directly provides a diagonal Gauss-Newton approximation that matches the cross-entropy loss, and is suitable for use with the aforementioned gradient. Together they are used for subsequent coordinate-by-coordinate scaling and robustness control.

[0042] To suppress mini-batch noise and reduce computational overhead, this embodiment applies time-dimension exponential moving average (EMA) and smooth updates to the gradient and curvature vector, respectively. The gradient is expressed in first-order momentum form: The curvature is updated using an EMA mechanism with a k-step interval: when First, the instantaneous curvature is obtained using the selected estimator. Press again Update; keep the rest of the steps. Unchanged. This utilizes EMA to filter out high-variance noise, thus... and It more robustly reflects the current training status; on the other hand, through the sparsely triggered curvature update strategy, the computation frequency of the diagonal Hessian is reduced from per step to per k steps, significantly controlling the additional overhead during training.

[0043] In summary, based on the above... The process of obtaining the full-parameter gradient, estimating the diagonal curvature based on Hutchinson or GNB, and smoothing the gradient and curvature using EMA, in step S105, completes the construction of key statistics for subsequent parameter updates without exceeding the range of the forward output and supervision signals. Provide directional information, It provides a curvature scale for each coordinate. Both can be calculated online in small batches, and the numerical stability is improved by exponential moving average, laying a data and statistical foundation for preconditioning and robust updates in subsequent steps.

[0044] In step S106, the gradient vector is scaled by the curvature vector according to the coordinates to form a preconditional gradient. The coordinate magnitude is clipped on the preconditional gradient. When the curvature vector is estimated to be zero, an update based on momentum sign is adopted, and weight decay is combined to suppress overfitting. Based on this, all trainable parameters of the detection model are updated.

[0045] When implementing the coordinate amplitude clipping, the process includes: comparing the amplitude of each coordinate update after curvature scaling with a threshold, compressing coordinate update amounts exceeding the threshold to within the threshold range; if the curvature vector of a certain coordinate is estimated to be zero or negative, then a sign update based on the historical momentum direction is used in the corresponding coordinate to maintain training stability, and a weight decay is applied to the parameters before each parameter update to limit the unbounded growth of the parameter amplitude.

[0046] In step S106, the momentum gradient vector and diagonal curvature vector obtained in the previous step are used as inputs. First, weight decay is applied to the parameters to suppress unbounded growth and overfitting. Then, curvature preconditioning and magnitude clipping are performed according to the coordinates to complete the parameter update. The standard update form is: first, weight decay is performed. Subsequently, a diagonal curvature precondition was applied and coordinate amplitude clipping was performed to obtain the updated result: .in, Let be the gradient vector obtained through momentum exponential moving average. To update the diagonal curvature vector at intervals and smoothed by exponential moving average, This is the scaling factor for the clipping threshold. The smallest positive number is set to avoid division by zero and negative curvature amplification; This indicates that the magnitude of the preconditional gradient for each coordinate is limited to a unit magnitude. This two-step update is performed sequentially within the same training step, ensuring that the initial weight decay is coordinated with the subsequent curvature-based adaptive step size.

[0047] The aforementioned coordinate-based curvature preconditioning essentially scales the update amounts in each dimension using local curvature: automatically reducing the update step size in dimensions with steeper curvature (larger Hessian) and appropriately increasing the step size in dimensions with flatter curvature (smaller Hessian), thereby weakening the influence of the condition number on the convergence speed and improving the progress efficiency in heterogeneous curvature landscapes. This dimension-adaptive property stems from the scaling effect of the diagonal Hessian preconditioner on the update amounts, matching the update size to the local geometry, suppressing oscillations and avoiding sluggish progress in flat directions.

[0048] The coordinate magnitude clipping step is used to handle curvature cases that are zero or near zero, or even estimated to be negative. This is because the preconditioning gradient includes... The denominator, when some coordinates If the value is too small or negative, it will cause the update amount of that coordinate to be abnormally amplified; at this time, The triggered coordinate-level clipping compresses the update magnitude to within a set threshold, while simultaneously degenerating into a robust momentum sign-based update in these dimensions. This means it only advances along the sign direction of historical momentum without relying on the problematic curvature scale, thus providing a reliable backoff strategy to maintain training stability and numerical safety. Threshold scaling factor Used to control the sensitivity of trigger clipping, preventing excessively large step sizes when the curvature estimate is too small or negative.

[0049] The above update process applies to all trainable parameter sets of the trunk, neck, and detection head along the parameter dimension, simultaneously completing curvature scaling, amplitude pruning, and sign backoff when necessary. Combined with pre-defined weight decay, the final update takes into account both structural regularization and curvature adaptability, enabling stable and efficient optimization in complex loss landscapes.

[0050] In step S107, the process of inputting the training samples into the detection model and updating all trainable parameters based on the precondition gradient is executed repeatedly until the convergence condition is met, and the optimized detection model is obtained.

[0051] When implementing step S107, an iterative training process is adopted: In each training step, the objective function and gradient are computed for the current mini-batch input, and two types of exponential moving averages are maintained—one for gradient momentum and the other for diagonal curvature. The diagonal curvature is not updated every step, but rather triggered at intervals, with an instantaneous estimate provided by the selected diagonal Hessian estimator followed by temporal smoothing. Parameter updates with weight decay and coordinate-wise pruning are then performed. This loop is organized from step 1 to step T. The core mechanism is to simultaneously maintain the EMA of the gradient and curvature, and update the diagonal Hessian only every k steps to control computational overhead and maintain numerical stability, avoiding excessively large updates when the curvature estimate is too small or unreliable. This process ensures that training progresses in a curvature-adaptive manner across complex loss scenarios.

[0052] To determine whether the iteration has reached the expected convergence state, performance and loss metrics can be collected at predetermined time points during training, and learning curves can be plotted to quantitatively evaluate the overall learning dynamics, convergence rate, and loss behavior. As an example, the changing trends of detection metrics (such as the comprehensive metric across the IoU threshold) and the loss of each component are statistically recorded at multiple training intervals to observe whether a plateau has been reached or whether significant improvement continues. If the validation curve stabilizes and no longer rises over a period of time, or the training-validation gap no longer narrows, the convergence condition can be determined to be met.

[0053] During loop execution, if some components of the curvature vector become negative or close to zero, the preconditioning gradient will automatically trigger coordinate-level clipping and degenerate into a robust update based on momentum sign in the corresponding dimension, thereby providing a reliable backoff strategy to prevent numerical instability; at the same time, by setting the control parameter of the clipping threshold, excessively large step sizes are avoided when curvature is unreliable.

[0054] In step S108, the optimized detection model is deployed on the image acquisition terminal or edge computing device to perform real-time inference on the input from the shooting scene and output standardized detection results containing the standardized category, location and confidence information for security management.

[0055] As an example, the following is one deployment method: During the deployment phase, the trained and optimized detection model is integrated into the image acquisition terminal or edge computing device as an executable inference component, enabling it to directly perform forward inference on video streams from cameras or periodically captured image frames. To meet the real-time constraints of continuous on-site monitoring, the selected model is based on the lightweight YOLOv11n platform and adopts an adapted backbone and detection head design. This allows it to maintain low inference latency and stable detection accuracy under limited computing power, making it suitable for on-site operation in resource-constrained environments such as camera-side or edge computing, thus avoiding the bandwidth and latency burden caused by transmitting large amounts of video back to a central server. After deployment, the system triggers inference according to the order of frame arrival, outputting the prediction results of candidate targets in each specification category frame by frame, maintaining continuity, and meeting the operational mode of simultaneous acquisition, judgment, and conclusion generation in security scenarios.

[0056] During inference, the model returns the class probability, bounding box location distribution, and relevant quality metrics for each candidate location, forming standardized detection results with a spatiotemporal index for subsequent recording and processing. To balance recall requirements and false alarm control in the field, the confidence information in the output is not a fixed threshold but is calibrated based on the application's risk tolerance: in situations where missing detections need to be prioritized, the threshold can be lowered to improve recall; in situations where alarm accuracy is emphasized, the threshold can be raised accordingly. This confidence threshold is explicitly defined as a detection confidence threshold at the deployment level and is not the same concept as the reliability percentage; therefore, in actual scenarios, the threshold will be calibrated and used strategically according to management requirements. The model maintains robust detection even in occluded, transparent material, and high-density human / object mixed scenarios, which is the foundation for the effectiveness of real-time edge deployment in complex construction sites.

[0057] To ensure continuous availability and closed-loop management after deployment, the standard detection results of each frame are structured and output as a triple containing standard category, location boundary, and confidence information at the edge. This can be expanded to include timestamps, collection point numbers, and image segment pointers, facilitating integration with on-duty / inspection logs, automatic generation of on-duty compliance statistics, or triggering of on-site alarms. Considering the diversity of construction scenarios, the deployment strategy differentiates between highly controlled workspaces and open, dynamically changing construction site environments: the former, with fixed workstations, controllable lighting, and standardized requirements, is suitable for continuous, full-coverage monitoring; the latter is more suitable for focused deployment in high-traffic paths, critical security zones, or temporarily sealed-off spaces to achieve higher on-site coverage effectiveness under limited computing power and viewing angle resources.

[0058] Based on the above deployment and operation methods, the optimized detection model can complete real-time inference of the shooting scene on edge devices, continuously outputting detection results with standardized categories, spatial locations and confidence information, providing credible compliance modules and early warning clues for security management; the balance between performance and efficiency and the enhancement of recall capability further meet the requirements of security scenarios that are sensitive to the cost of missed detection, thereby improving the continuity and operability of compliance supervision without increasing additional communication and computing power costs.

[0059] Furthermore, the various English abbreviations in the above embodiments are explained as follows: Based on the same line of thought, such as Figure 2 The diagram shown is a structural block diagram of a video data standardization review system based on artificial intelligence, provided in an embodiment of the present invention. The system includes: The acquisition module 201 is used to acquire images or video frames from the shooting scene, label the targets according to the standard categories, and obtain training samples. The model building module 202 is used to build a detection model based on SC-YOLO. The detection model includes: feature extraction using the CSPDarknet backbone network, in which C3k2 and SPPF modules are set, and a C2PSA attention module is set at the tail of the CSPDarknet backbone network; upsampling, connecting, and fusing multi-scale features using a neck structure of feature pyramid and path aggregation; and outputting class probability, bounding box position distribution, and box quality index using a multi-scale detection head, in which depthwise separable convolution is used to reduce computational cost. The feature extraction module 203 is used to input the labeled training samples into the detection model, extract multi-level semantic and texture features through the CSPDarknet backbone network, perform multi-scale semantic fusion through the neck structure, and input the samples into the multi-scale detection head, outputting the class probability and bounding box distribution of each candidate position. The comprehensive loss construction module 204 is used to construct a weighted comprehensive loss including bounding box regression loss, classification loss and distributed focus loss based on the category probability, bounding box distribution and annotation results of the forward output, so as to simultaneously constrain the positioning accuracy, category determination and bounding box quality. The second-order estimation module 205 is used to obtain the gradient vector of the weighted comprehensive loss for all trainable parameters of the detection model, and to estimate the local curvature of the objective function formed by the weighted comprehensive loss. The curvature estimation includes: obtaining the diagonal curvature vector using a diagonal Hessian estimator formed by the inner product of a random vector and the gradient or a Gauss-Newton-Bartlett estimator based on the model's log odds output and sampling labels, and performing exponential moving averages on the gradient vector and the curvature vector in the time dimension to suppress noise. The parameter update module 206 is used to scale the gradient vector with the curvature vector according to the coordinates to form a precondition gradient, perform coordinate magnitude clipping on the precondition gradient to limit anomalous steps, and degenerate into a momentum sign-based update when the curvature vector estimate is zero or unreliable, and combine it with weight decay to suppress overfitting, thereby updating all trainable parameters of the detection model. The iterative training module 207 is used to repeatedly execute the process of inputting the training samples into the detection model and updating all trainable parameters based on the precondition gradient until the convergence condition is met, thereby obtaining the optimized detection model. The deployment module 208 is used to deploy the optimized detection model on the image acquisition terminal or edge computing device, perform real-time inference on the input at the shooting site, and output the standardized detection results containing the standardized category, location and confidence information for security management.

[0060] The specific details of the above system have been described in detail in the method section of the implementation plan. For any undisclosed details, please refer to the implementation plan of the method section, and therefore will not be repeated here.

[0061] This system achieves the following: more robust detection, with multi-scale feature fusion and C2PSA attention helping to maintain the recognition and localization of small-sized and transparent PPEs in crowded, occluded, and complex backgrounds, and its recall capability is more suitable for safety and compliance scenarios; more reliable training, with Sophia's Hessian diagonal estimation and element-level pruning providing curvature-adaptive updates and degradation protection, mitigating gradient oscillations and accelerating stable convergence; more efficient deployment, with the detection head employing depthwise separable convolution combined with a lightweight backbone, balancing inference speed and accuracy, suitable for real-time review at resource-constrained edge environments; and more practical generalization, maintaining consistent performance trends across different datasets and diverse field conditions, supporting scenario migration from fixed factory areas to dynamic construction sites, thereby providing continuous, quantitative, and traceable regulatory reviews for safety management.

[0062] The accompanying drawings are merely illustrative of the processes included in the methods according to exemplary embodiments of this disclosure and are not intended to be limiting. It is readily understood that the processes shown in the drawings do not indicate or limit the temporal order of these processes. Furthermore, it is readily understood that these processes may be executed synchronously or asynchronously, for example, in multiple modules.

[0063] It should be noted that although several modules or units of the system have been mentioned in the detailed description above, this division is not mandatory. In fact, according to exemplary embodiments of this disclosure, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0064] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the claims.

[0065] It should be understood that this disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims.

Claims

1. A video data normative review method based on artificial intelligence, characterized in that, The method includes: Collect images or video frames from the shooting scene, label the targets according to the standardized categories, and obtain training samples; A detection model based on SC-YOLO is constructed. The detection model includes: feature extraction using the CSPDarknet backbone network, in which C3k2 and SPPF modules are set, and a C2PSA attention module is set at the tail of the CSPDarknet backbone network; upsampling, connecting, and fusing multi-scale features using a neck structure of feature pyramid and path aggregation; and outputting class probability, bounding box position distribution, and box quality index using a multi-scale detection head. The multi-scale detection head uses depthwise separable convolution to reduce computational cost. The labeled training samples are input into the detection model, and multi-level semantic and texture features are extracted by the CSPDarknet backbone network. After multi-scale semantic fusion is performed by the neck structure, the samples are input into the multi-scale detection head, and the class probability and bounding box distribution of each candidate position are output. Based on the category probabilities, bounding box distribution, and annotation results of the forward output, a weighted comprehensive loss is constructed, which includes bounding box regression loss, classification loss, and distributed focus loss, to simultaneously constrain positioning accuracy, category determination, and bounding box quality. The gradient vector of the weighted comprehensive loss is obtained for all trainable parameters of the detection model, and the local curvature of the objective function formed by the weighted comprehensive loss is estimated. The curvature estimation includes: obtaining the diagonal curvature vector using a diagonal Hessian estimator formed by the inner product of a random vector and the gradient or a Gauss-Newton-Bartlett estimator based on the model's log odds output and sampling labels, and performing exponential moving averages on the gradient vector and the curvature vector in the time dimension to suppress noise. The gradient vector is scaled by the curvature vector to form a preconditional gradient according to the coordinates. The coordinate magnitude is clipped on the preconditional gradient. When the curvature vector is estimated to be zero, an update based on momentum sign is adopted, and weight decay is combined to suppress overfitting. Based on this, all trainable parameters of the detection model are updated. The process of inputting the training samples into the detection model and updating all trainable parameters based on the precondition gradient is repeated until the convergence condition is met, thus obtaining the optimized detection model. The optimized detection model is deployed on an image acquisition terminal or edge computing device to perform real-time inference on the input from the shooting scene and output standardized detection results containing the standardized category, location and confidence information for security management.

2. The video data standardization review method based on artificial intelligence according to claim 1, characterized in that, The CSPDarknet backbone network employs cross-stage partial connections to divide the feature mapping into a preservation path and a transformation path, which are then merged in subsequent stages to enhance gradient flow and reduce redundant computation; the C3k2 module is used to improve feature representation efficiency while maintaining the receptive field. The SPPF module is used to aggregate multi-scale contextual information; the C2PSA attention module is used to suppress background interference and highlight salient regions related to canonical categories.

3. The video data standardization review method based on artificial intelligence according to claim 1, characterized in that, The neck structure, based on feature pyramids and path aggregation, is executed sequentially: High-level semantic features are upsampled, and the upsampled results are connected and fused with low-level detail features of adjacent levels. The C3k2 module is then used for nonlinear recombination to form multi-scale fused features that include both semantics and details.

4. The video data standardization review method based on artificial intelligence according to claim 1, characterized in that, The multi-scale detection head feeds the multi-scale fused features output from the neck structure into parallel prediction branches. These prediction branches employ depthwise separable convolutions to reduce computational load and output the following: The probability distribution used for class determination, the center and boundary distribution used for bounding box localization, and the score used for box quality measurement are used to adapt to the standard categories of different sizes.

5. The video data standardization review method based on artificial intelligence according to claim 1, characterized in that, The weighted comprehensive loss is calculated by first calculating the bounding box regression loss to penalize the positional difference between the predicted box and the labeled box, then calculating the classification loss to measure the consistency between the class probability and the true class, and simultaneously calculating the distributed focus loss to improve the fineness of the boundary distribution. Finally, the bounding box regression loss, the classification loss and the distributed focus loss are linearly combined with preset weights to obtain the objective function used for backpropagation.

6. The video data standardization review method based on artificial intelligence according to claim 1, characterized in that, When performing the coordinate magnitude clipping, the following are included: The magnitude of each coordinate update after curvature scaling is compared with the threshold, and the coordinate update exceeding the threshold is compressed to the threshold range. If the curvature vector of a certain coordinate is estimated to be zero or negative, the sign update based on the historical momentum direction is used in the corresponding coordinate to maintain training stability, and the weight decay is applied to the parameters before each parameter update to limit the unbounded growth of the parameter magnitude.

7. An artificial intelligence-based video data standardization review system, the system comprising: The acquisition module is used to acquire images or video frames from the shooting scene, label the targets according to the standard categories, and obtain training samples; The model building module is used to construct a detection model based on SC-YOLO. The detection model includes: feature extraction using the CSPDarknet backbone network, which includes a C3k2 module and an SPPF module, and a C2PSA attention module at the tail of the CSPDarknet backbone network; upsampling, connecting, and fusing multi-scale features using a neck structure of feature pyramids and path aggregation; and outputting class probabilities, bounding box position distributions, and box quality indices using a multi-scale detection head, wherein depthwise separable convolutions are used in the multi-scale detection head to reduce computational cost. The feature extraction module is used to input the labeled training samples into the detection model, extract multi-level semantic and texture features through the CSPDarknet backbone network, perform multi-scale semantic fusion through the neck structure, and input the samples into the multi-scale detection head, outputting the class probability and bounding box distribution of each candidate position. The comprehensive loss construction module is used to construct a weighted comprehensive loss including bounding box regression loss, classification loss and distributed focus loss based on the category probability, bounding box distribution and annotation results of the forward output, so as to simultaneously constrain the positioning accuracy, category determination and bounding box quality. The second-order estimation module is used to obtain the gradient vector of the weighted comprehensive loss for all trainable parameters of the detection model, and to estimate the local curvature of the objective function formed by the weighted comprehensive loss. The curvature estimation includes: obtaining the diagonal curvature vector using a diagonal Hessian estimator formed by the inner product of a random vector and the gradient or a Gauss-Newton-Bartlett estimator based on the model's log odds output and sampling labels, and performing exponential moving averages in the time dimension on the gradient vector and the curvature vector respectively to suppress noise. The parameter update module is used to scale the gradient vector by the curvature vector to form a preconditional gradient, perform coordinate magnitude clipping on the preconditional gradient to limit anomalous steps, and degenerate into a momentum sign-based update when the curvature vector estimate is zero or unreliable, and combine it with weight decay to suppress overfitting, thereby updating all trainable parameters of the detection model. The iterative training module is used to repeatedly execute the process of inputting the training samples into the detection model and updating all trainable parameters based on the precondition gradient until the convergence condition is met, thereby obtaining the optimized detection model. The deployment module is used to deploy the optimized detection model on the image acquisition terminal or edge computing device, perform real-time inference on the input from the shooting scene, and output standardized detection results containing the standardized category, location and confidence information for security management.

Images