An optimization method and system for embedded image recognition algorithm

Abstract

The present invention relates to the field of data processing technology, and in particular to an optimization method and system for embedded image recognition algorithms. The present invention obtains risk assessment feature data by deeply mining multimodal environmental data, and generates dynamic risk levels in combination with scene feature data. The initial image processing parameters are then obtained in combination with hardware status data. This method of dynamically configuring initial parameters based on the actual hardware capabilities and risk requirements can ensure fire identification accuracy while meeting real-time requirements, alleviate the contradiction between high resolution and limited computing power, and then obtain real-time matching values ​​based on computing performance data, risk assessment feature data, and initial image processing parameters to evaluate whether hardware resources under the current parameter configuration can meet the algorithm processing requirements. When the computing delay does not meet the real-time requirements, the initial image processing parameters are dynamically adjusted to ensure the reliable operation of the embedded image recognition algorithm in the fire warning scenario.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to an optimization method and system for embedded image recognition algorithm. BACKGROUND

[0002] The embedded image recognition algorithm refers to the image recognition technology deployed on embedded devices (such as single-chip microcomputers, ARM processors, FPGA / ASIC chips, etc. terminal devices with limited computing resources), which realizes the classification, detection, segmentation, etc. of image content by directly processing, analyzing and feature extraction on the collected image data locally on the device. The current embedded image recognition algorithm has been widely applied in fire warning field, mainly relying on the edge computing capability of embedded system, through the image processing algorithm deployed in the front-end device (such as camera, sensor node), the image data in the monitoring scene is analyzed in real time, so as to realize the early detection and warning of fire.

[0003] In the scene of applying embedded image recognition algorithm in fire warning, the contradiction between computing power and real-time is a key bottleneck problem, which mainly reflects the conflict between high-resolution image processing demand and limited computing resources of embedded device, and the mismatch between computing delay of complex model and real-time requirement of actual scene. The conflict of the two aspects makes it difficult for embedded image recognition algorithm to balance accuracy and timeliness in fire warning. SUMMARY

[0004] The main purpose of the present application is to provide an optimization method and system for embedded image recognition algorithm, which aims to solve the technical problems in the prior art.

[0005] The present application provides an optimization method for embedded image recognition algorithm, comprising:

[0006] Obtaining multi-modal monitoring data monitored by a sensor, wherein the multi-modal monitoring data comprises scene feature data, multi-modal environment data and hardware state data;

[0007] Obtaining risk assessment feature data according to the multi-modal environment data;

[0008] Obtaining dynamic risk level according to the risk assessment feature data and the scene feature data;

[0009] Obtaining computing performance data and running performance data according to the hardware state data, and obtaining initial image processing parameters according to the computing performance data, running performance data and dynamic risk level;

[0010] Obtaining real-time matching value according to the computing performance data, risk assessment feature data and initial image processing parameters;

[0011] determining whether the real-time matching value is greater than a preset threshold value;

[0012] if the real-time matching value is greater than the preset threshold value, determining that the calculation delay meets the real-time requirement;

[0013] if the real-time matching value is not greater than the preset threshold value, determining that the calculation delay does not meet the real-time requirement, and adjusting the initial image processing parameter according to the real-time matching value and the dynamic risk level until the real-time matching value is greater than the preset threshold value.

[0014] Preferably, the step of obtaining risk assessment feature data according to the multi-modal environment data comprises:

[0015] spectrum image data, thermal radiation image data and gaseous molecule data are obtained according to the multi-modal environment data, and spectrum alignment images and thermal radiation alignment images are obtained according to the spectrum image data and the thermal radiation image data;

[0016] global chroma variance is obtained according to the spectrum alignment images, and a spectrum feature vector is obtained by principal component analysis according to the global chroma variance;

[0017] timestamp feature points are obtained according to the spectrum alignment images and the thermal radiation alignment images, and time alignment gas data are obtained according to the timestamp feature points and the gaseous molecule data;

[0018] a CO concentration sequence and a concentration gradient are obtained according to the time alignment gas data, and a color feature vector is obtained according to the CO concentration sequence, the concentration gradient and the spectrum feature vector;

[0019] high-temperature centroid coordinates and a temperature gradient matrix are obtained according to the thermal radiation alignment images, and a spatio-temporal adjacency matrix is obtained according to the high-temperature centroid coordinates;

[0020] a thermodynamic gradient tensor is obtained according to the spatio-temporal adjacency matrix and the temperature gradient matrix, and a dynamic texture entropy value matrix is obtained according to the spatio-temporal adjacency matrix.

[0021] Preferably, the step of obtaining a dynamic risk level according to the risk assessment feature data and the scene feature data comprises:

[0022] spatial positioning data, optical feature parameters and structure topology parameters are obtained according to the scene feature data, and a regional function type is obtained according to the spatial positioning data and the structure topology parameters;

[0023] a material risk level is obtained according to the regional function type and the optical feature parameters, and a spatial attribute category is obtained according to the regional function type, the material risk level and the structure topology parameters.

[0024] correlation analysis is performed on the color feature vector, dynamic texture entropy value matrix and thermodynamic gradient tensor respectively with the spatial attribute category to obtain corresponding modal correlation coefficients, and a corresponding correlation weight index is obtained according to the modal correlation coefficients;

[0025] A normalized feature data set is obtained according to the color feature vector, dynamic texture entropy value matrix and thermodynamic gradient tensor, and a dynamic risk level is obtained according to the normalized feature data set and the correlation weight index.

[0026] Preferably, the step of obtaining initial image processing parameters according to the computing performance data, running performance data and dynamic risk level comprises:

[0027] A core frequency, a benchmark frequency, an architecture coefficient, a utilization margin, a remaining memory capacity and a memory fragmentation rate are obtained according to the computing performance data, and a resource comprehensive index is obtained according to the core frequency, benchmark frequency, architecture coefficient and utilization margin;

[0028] A memory health degree is obtained according to the remaining memory capacity and the memory fragmentation rate;

[0029] A real-time running temperature, a safety threshold temperature and a maximum allowable temperature are obtained according to the running performance data, and a computing health degree is obtained according to the real-time running temperature, safety threshold temperature and maximum allowable temperature;

[0030] A resolution parameter is obtained according to the resource comprehensive index and the computing health degree;

[0031] A risk-resource mapping table is obtained, and a computing demand benchmark and a memory demand benchmark are obtained according to the risk-resource mapping table and the dynamic risk level;

[0032] A model architecture parameter is obtained according to the memory demand benchmark and the memory health degree, and a computing precision parameter is obtained according to the resource comprehensive index and the computing demand benchmark.

[0033] Preferably, the step of obtaining real-time matching values according to the computing performance data, risk assessment feature data and initial image processing parameters comprises:

[0034] A hardware floating-point computing power effective value, a chip junction temperature, a cycle operation number, a memory bandwidth and a processing period are obtained according to the computing performance data, and a CPU real-time computing power is obtained according to the hardware floating-point computing power effective value, core frequency, cycle operation number, chip junction temperature and safety threshold temperature;

[0035] Processing demand data is obtained according to the risk assessment feature data, and a processing time threshold and an algorithm calculation amount are obtained according to the processing demand data.

[0036] obtain a theoretical processing time according to the algorithm calculation amount and the CPU real-time computing power, and obtain an algorithm matching degree according to the theoretical processing time and the processing time threshold value;

[0037] obtain a memory available space according to the remaining memory capacity, the memory fragmentation rate, the memory bandwidth and the processing period;

[0038] obtain a memory requirement space and a memory requirement bandwidth according to the resolution parameter, the model architecture parameter and the calculation precision parameter, and obtain a memory matching degree according to the memory requirement space, the memory available space, the memory bandwidth and the memory requirement bandwidth;

[0039] obtain a real-time matching value according to the algorithm matching degree and the memory matching degree.

[0040] Preferably, the step of adjusting the initial image processing parameter according to the real-time matching value and the dynamic risk level comprises:

[0041] obtain an adjustment intensity level according to the real-time matching value and the dynamic risk level, and obtain an adjustment constraint condition according to the dynamic risk level;

[0042] obtain a parameter adjustment data set according to the adjustment intensity level and the adjustment constraint condition, wherein the parameter adjustment data set comprises a plurality of resolution adjustment data, model architecture adjustment data and calculation precision adjustment data;

[0043] arrange and combine the plurality of resolution adjustment data, model architecture adjustment data and calculation precision adjustment data to obtain a plurality of parameter adjustment combination sets;

[0044] screen the plurality of parameter adjustment combination sets according to the real-time matching value and the adjustment constraint condition to obtain an optimal parameter adjustment combination, and return the optimal parameter adjustment combination to the step of obtaining a real-time matching value according to the calculation performance data, risk assessment feature data and initial image processing parameter as the initial image processing parameter.

[0045] The application further provides an optimization system for an embedded image recognition algorithm, comprising:

[0046] a first obtaining module configured to obtain multi-modal monitoring data monitored by a sensor, wherein the multi-modal monitoring data comprises scene feature data, multi-modal environment data and hardware state data;

[0047] a second obtaining module configured to obtain risk assessment feature data according to the multi-modal environment data;

[0048] a third obtaining module, configured to obtain a dynamic risk level according to the risk assessment feature data and the scene feature data;

[0049] a fourth obtaining module, configured to obtain computing performance data and running performance data according to the hardware state data, and obtain initial image processing parameters according to the computing performance data, the running performance data and the dynamic risk level;

[0050] a fifth obtaining module, configured to obtain a real-time matching value according to the computing performance data, the risk assessment feature data and the initial image processing parameters;

[0051] a judging module, configured to judge whether the real-time matching value is greater than a preset threshold value;

[0052] if the real-time matching value is greater than the preset threshold value, it is determined that the computing delay meets the real-time requirement;

[0053] if the real-time matching value is not greater than the preset threshold value, it is determined that the computing delay does not meet the real-time requirement, and the initial image processing parameters are adjusted according to the real-time matching value and the dynamic risk level until the real-time matching value is greater than the preset threshold value.

[0054] Preferably, the third obtaining module comprises:

[0055] a first obtaining unit, configured to obtain spatial positioning data, optical feature parameters and structure topology parameters according to the scene feature data, and obtain a region function type according to the spatial positioning data and the structure topology parameters;

[0056] a second obtaining unit, configured to obtain a material risk level according to the region function type and the optical feature parameters, and obtain a spatial attribute category according to the region function type, the material risk level and the structure topology parameters;

[0057] a third obtaining unit, configured to perform correlation analysis on the color feature vector, the dynamic texture entropy value matrix and the thermodynamic gradient tensor respectively and the spatial attribute category, to obtain corresponding modal correlation coefficients, and obtain corresponding correlation weight indexes according to the modal correlation coefficients;

[0058] a fourth obtaining unit, configured to obtain a normalized feature data set according to the color feature vector, the dynamic texture entropy value matrix and the thermodynamic gradient tensor, and obtain a dynamic risk level according to the normalized feature data set and the correlation weight indexes.

[0059] The application further provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above-mentioned optimization method for embedded image recognition algorithm when executing the computer program.

[0060] The application further provides a computer readable storage medium, which stores a computer program, and the computer program realizes the steps of the optimization method for the embedded image recognition algorithm when executed by a processor.

[0061] The application has the advantages that: the application collects multi-modal data including scene feature data, multi-modal environment data and hardware state data from different types of sensors, provides a comprehensive data basis for subsequent analysis, avoids the limitations of single data sources, then obtains risk assessment feature data by deeply mining the multi-modal environment data, can more accurately identify fire signs, can detect weak flames and temperature changes in the early stage of fire, gains time for subsequent processing, reduces the real-time risk caused by complex model calculation delay, then further combines the scene feature data to generate a dynamic risk level, and then combines the calculation performance data, the running performance data and the dynamic risk level to obtain initial image processing parameters, this method of dynamically configuring initial parameters according to the actual ability of hardware and risk requirements can meet the real-time requirements while ensuring the fire identification accuracy, and can alleviate the contradiction between high resolution and limited computing power, then the real-time matching value is obtained according to the calculation performance data, the risk assessment feature data and the initial image processing parameters, so as to evaluate whether the hardware resources can meet the algorithm processing requirements under the current parameter configuration, this quantitative evaluation method enables the system to dynamically adjust according to the actual situation, avoids the waste of computing power or insufficient real-time caused by blind parameter configuration, when the calculation delay does not meet the real-time requirement, the initial image processing parameters are dynamically adjusted according to the real-time matching value and the dynamic risk level until the real-time requirement is met, this closed-loop adjustment mechanism continuously optimizes the parameter configuration, can effectively solve the contradiction between computing power and real-time, and ensures the reliable operation of the embedded image recognition algorithm in the fire warning scene. BRIEF DESCRIPTION OF DRAWINGS

[0062] Figure 1 The figure is a method flowchart of an embodiment of the application.

[0063] Figure 2 The figure is a system structure diagram of an embodiment of the application.

[0064] Figure 3 The figure is a computer device internal structure diagram of an embodiment of the application.

[0065] The implementation, functional features and advantages of the application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0066] It should be understood that the specific embodiments described herein are merely intended to explain the application, and are not intended to limit the application.

[0067] As Figure 1As shown, the application provides an optimization method for embedded image recognition algorithm, comprising:

[0068] S1, acquiring multi-modal monitoring data monitored by a sensor, wherein the multi-modal monitoring data comprises scene feature data, multi-modal environment data and hardware state data;

[0069] S2, acquiring risk assessment feature data according to the multi-modal environment data;

[0070] S3, acquiring dynamic risk level according to the risk assessment feature data and the scene feature data;

[0071] S4, acquiring computing performance data and running performance data according to the hardware state data, and acquiring initial image processing parameters according to the computing performance data, running performance data and dynamic risk level;

[0072] S5, acquiring real-time matching value according to the computing performance data, risk assessment feature data and initial image processing parameters;

[0073] S6, judging whether the real-time matching value is greater than a preset threshold value;

[0074] If the real-time matching value is greater than the preset threshold value, it is determined that the calculation delay meets the real-time requirement;

[0075] If the real-time matching value is not greater than the preset threshold value, it is determined that the calculation delay does not meet the real-time requirement, and the initial image processing parameters are adjusted according to the real-time matching value and the dynamic risk level until the real-time matching value is greater than the preset threshold value.

[0076] As described in steps S1-S6 above, the present application collects multi-modal data from different types of sensors, where multi-modal data refers to a collection of data reflecting different dimensions of information about the monitoring target, including scene feature data, multi-modal environment data, and hardware state data. Scene feature data refers to image or video data that directly represents the visual characteristics of the monitoring scene. Multi-modal environment data refers to environmental physical and chemical characteristic data collected by non-visual sensors. Hardware state data refers to the state parameters of the embedded device hardware during runtime. These data provide a comprehensive data foundation for subsequent analysis, avoiding the limitations of single data sources. Then, feature extraction and analysis are performed on the multi-modal environment data to obtain risk assessment feature data, which refers to a set of key features extracted from the multi-modal environment data for fire risk assessment, including spectral feature vectors, color feature vectors, and thermodynamic gradient tensors. By deeply mining data features, fire signs can be more accurately identified. For example, using the color features (global chroma variance, spectral feature vectors) of spectral image data and the temperature distribution (temperature gradient matrix, thermodynamic gradient tensor) of thermal radiation image data, weak flames and temperature changes can be detected in the early stages of a fire. Compared with traditional methods that rely only on single image features, fires can be detected earlier, providing time for subsequent processing and reducing the real-time risk caused by complex model calculation delays. Then, the scene feature data is further combined to generate a dynamic risk level, which refers to the real-time calculation of fire risk levels (such as low, medium, and high risk) based on risk assessment feature data and scene feature data. The dynamic risk level provides priority guidance for subsequent parameter adjustment. For example, in high-risk scenarios, real-time requirements are higher, and the system can prioritize allocating computing resources, using higher resolution and more complex models to ensure the accuracy of fire identification.In a low-risk scenario, the resource consumption can be appropriately reduced, a lightweight model and a lower resolution can be adopted, the calculation delay and the resource utilization can be balanced, the excessive occupation of limited computing power by high-resolution image processing can be avoided, then the computing performance data and the running performance data are extracted from the hardware state data, the computing performance data refers to static parameters representing the hardware computing capability of the embedded device, such as the number of cores, the main frequency, the computing power peak value, etc., the running performance data refers to real-time state parameters of the embedded device in running, and the initial image processing parameters are obtained in combination with the dynamic risk level, wherein the initial image processing parameters refer to a set of image processing algorithm parameters pre-configured based on the hardware performance and the risk level, including the resolution parameter, the model architecture parameter and the calculation precision parameter, this method of dynamically configuring the initial parameters according to the actual capability of the hardware and the risk demand can guarantee the fire identification accuracy while meeting the real-time requirement, and can relieve the contradiction between the high resolution and the limited computing power, then the real-time matching value is obtained according to the computing performance data, the risk assessment feature data and the initial image processing parameters, the real-time matching value is a comprehensive index for measuring the matching degree of the current image processing parameters and the hardware performance, the real-time matching value is compared with the preset threshold value to determine whether the current calculation delay meets the real-time requirement, if the real-time matching value is greater than the threshold value, it means that the calculation delay is within the acceptable range, the system continues to run with the current configuration, if the real-time matching value is not greater than the threshold value, it means that the hardware resources cannot meet the algorithm processing demand under the current parameter configuration, there may be a problem of high calculation delay, and the parameter needs to be adjusted; this quantitative evaluation method enables the system to dynamically adjust according to the actual situation, avoids the waste of computing power or the lack of real-time performance caused by blind parameter configuration, when the calculation delay does not meet the real-time requirement, the initial image processing parameters are dynamically adjusted (such as reducing the resolution, switching to a lightweight model, adjusting the calculation precision) according to the real-time matching value and the dynamic risk level, and the real-time matching value is recalculated until the real-time requirement is met, this closed-loop adjustment mechanism continuously optimizes the parameter configuration, can effectively solve the contradiction between the computing power and the real-time performance, and ensures the reliable operation of the embedded image recognition algorithm in the fire warning scene.

[0077] In one embodiment, the step S2 of obtaining risk assessment feature data according to the multi-modal environment data comprises:

[0078] S21, obtaining spectral image data, thermal radiation image data and gaseous molecule data according to the multi-modal environment data, and obtaining a spectral alignment image and a thermal radiation alignment image according to the spectral image data and the thermal radiation image data;

[0079] S22, obtaining a global chroma variance according to the spectral alignment image, and obtaining a spectral feature vector according to the global chroma variance through principal component analysis;

[0080] S23, acquiring timestamp feature points according to the spectrum alignment image and the thermal radiation alignment image, and acquiring time alignment gas data according to the timestamp feature points and the gaseous molecule data;

[0081] S24, acquiring CO concentration sequence and concentration gradient according to the time alignment gas data, and acquiring color feature vector according to the CO concentration sequence, concentration gradient and spectrum feature vector;

[0082] S25, acquiring high temperature centroid coordinates and temperature gradient matrix according to the thermal radiation alignment image, and acquiring space-time adjacency matrix according to the high temperature centroid coordinates;

[0083] S26, acquiring thermodynamic gradient tensor according to the space-time adjacency matrix and the temperature gradient matrix, and acquiring dynamic texture entropy value matrix according to the space-time adjacency matrix.

[0084] As described in steps S21-S26 above, in the present application, the fire risk features in the environment are extracted from multiple dimensions by obtaining spectral image data, thermal radiation image data and gaseous molecule data. The spectral image data refers to the scene image data collected by a spectral sensor (such as a multispectral / hyperspectral camera), which records the radiation intensity distribution at different wavelengths (such as visible light, near-infrared). The thermal radiation image data refers to the scene temperature distribution image collected by a thermal imager, which reflects the thermal radiation intensity of the object surface (usually displayed in grayscale or pseudo-color to show the temperature level). The gaseous molecule data refers to the gaseous molecule concentration data in the environment collected by a gas sensor (such as an infrared gas detector, an electrochemical sensor), such as the real-time concentration values of fire-related gases such as CO, CO2, smoke particles, etc. Then, the feature points in the spectral image data and the thermal radiation image data are extracted, and the mis-matched points are removed by RANSAC to align the visible light and infrared images, obtaining the spectral aligned image and the thermal radiation aligned image. Then, the spectral aligned image is converted from RGB to HSV color space, separating the hue (H) and saturation (S) channels, and calculating the global variance of the H channel and the S channel respectively to obtain the global hue variance, which refers to the variance of the hue value (such as the saturation and hue in the HSV space) of all pixels in the spectral image, used to measure the dispersion degree of the overall color distribution of the image. Then, principal component analysis is performed on the global hue variance to further obtain the spectral feature vector, which refers to the feature vector obtained by dimensionality reduction on the spectral features such as global hue variance. This method of in-depth mining of color features in spectral images of fire scenes can more accurately capture these features, thereby improving the accuracy of the use of spectral image information in fire risk assessment, solving the shortcomings of the prior art in spectral color feature extraction. Next, the pyramid LK optical flow method is used to track the FAST corner points in the spectral aligned image and the thermal radiation aligned image, generating time-stamped trajectories, and retaining points with motion speed > 5 pixels / frame and trajectory length > 5 frames as time-stamped feature points. The time-stamped feature points refer to spatial feature points with timestamps extracted by feature point detection algorithms (such as SIFT, ORB) in multiple frames of spectral images and thermal radiation images. Then, the linear interpolation method is used to synchronize the gaseous molecule data to the timestamp of each image frame to generate time-aligned gas data. In a fire scene, the generation and change of gas are related to the spectral and thermal radiation changes of the flame in time. The prior art may not strictly align the timestamps of multi-modal data, resulting in timing deviation during feature fusion. The present scheme deeply correlates different modal data in time and spatial dimensions, taking into account the consistency and relevance of different physical phenomena (optics, heat, gas change) in time when a fire occurs, which can more comprehensively understand the physical and chemical changes during the development of a fire, improving the accuracy and comprehensiveness of risk assessment. Then, the CO concentration sequence and concentration gradient are obtained according to the time-aligned gas data.The CO concentration sequence refers to a sequence of CO concentration values arranged in time sequence, and the concentration gradient refers to the rate of change of CO concentration in space. Since the CO concentration gradient can reflect the speed and trend of fire development, it is combined with the spectral feature vector to obtain a color feature vector. The color feature vector refers to a multi-dimensional feature vector formed by fusing the CO concentration sequence, the concentration gradient, and the spectral feature vector, which can more comprehensively describe the characteristics of the fire scene. Then, the thermal radiation alignment image is subjected to threshold segmentation, and the area with a temperature greater than ambient temperature + 30°C is extracted as a high-temperature area. Each area is labeled using connected component analysis to obtain high-temperature centroid coordinates, which refer to the centroid (geometric center) coordinates of the high-temperature area in the thermal radiation image, used to locate the spatial position of the heat source. Then, a 3x3 Gaussian derivative kernel is used to calculate the gradient matrix of the high-temperature area to obtain a temperature gradient matrix, which refers to a matrix composed of temperature gradient values of each pixel point in the thermal radiation image. Each element represents the temperature change rate (such as gradient amplitude and direction) of the point in the x and y directions. Then, each high-temperature centroid coordinate is regarded as a node, and the Euclidean distance between the nodes is calculated. If the distance is less than 50 pixels (empirical threshold), an undirected edge is established, and the weight is the absolute value of the temperature difference. The high-temperature centroids of adjacent frames are matched using the Hungarian algorithm. If the matching is successful, a time edge is established, and the weight is the centroid displacement speed. Thus, a space-time adjacency matrix is obtained, which refers to a matrix constructed based on the space-time correlation of the high-temperature centroid coordinates, representing the spatial adjacency relationship of the high-temperature areas in adjacent time frames. Through this method, the propagation, diffusion, and dynamic changes of thermal radiation in the fire can be more accurately understood. Then, the temperature gradient matrix and the space-time adjacency matrix are subjected to tensor product operation to generate a thermodynamic gradient tensor, which refers to a multi-dimensional physical quantity describing the spatial temperature gradient and the time adjacency relationship, thus more accurately representing the thermodynamic behavior of the fire source. Finally, for each high-temperature area, the local binary pattern texture feature is calculated to generate a dynamic texture entropy value matrix, which refers to a matrix formed by calculating the entropy value of the texture feature (such as the flicker of the flame and the turbulence of the smoke) of the thermal radiation image sequence. It can be used as a key feature to distinguish between "dynamic flame" and "static high temperature".

[0085] In one embodiment, the step S3 of obtaining a dynamic risk level according to the risk assessment feature data and the scene feature data comprises:

[0086] S31, obtaining spatial positioning data, optical feature parameters and structure topology parameters according to the scene feature data, and obtaining a regional function type according to the spatial positioning data and the structure topology parameters;

[0087] S32, obtaining a material risk level according to the region function type and the optical characteristic parameter, and obtaining a space attribute category according to the region function type, the material risk level and the structure topology parameter;

[0088] S33, performing correlation analysis on the color feature vector, the dynamic texture entropy value matrix and the thermodynamic gradient tensor respectively with the space attribute category to obtain corresponding modal correlation coefficients, and obtaining corresponding correlation weight indexes according to the modal correlation coefficients;

[0089] S34, obtaining a normalized feature data set according to the color feature vector, the dynamic texture entropy value matrix and the thermodynamic gradient tensor, and obtaining a dynamic risk level according to the normalized feature data set and the correlation weight indexes.

[0090] As described in steps S31-S34, the present application provides a three-dimensional environmental awareness capability for a scene-aware fire monitoring system by acquiring spatial positioning data, optical feature parameters, and structural topology parameters. The spatial positioning data refers to data describing the position, size, and geometric relationship of objects or regions in a scene in three-dimensional space. The optical feature parameters refer to visual features extracted from images or spectra, reflecting the physical properties of object surfaces. The structural topology parameters refer to data describing the spatial connection relationship and network structure of objects in a scene. Traditional fire monitoring systems use a "one-size-fits-all" algorithm, using a fixed feature set (such as a uniform monitoring temperature threshold, smoke concentration) for all scenes, ignoring the differences in fire characteristics of different scenes. For example, the key features for monitoring forest fires are flame spectrum, heat radiation diffusion speed, and wind direction, while the key features for monitoring shopping mall fires are smoke particle concentration, personnel density change, and electrical equipment abnormalities. Therefore, existing systems cannot automatically switch monitoring strategies, resulting in high forest fire miss rate (relying on smoke identification) and high shopping mall false alarm rate (air conditioner hot air triggering temperature threshold). Therefore, the present application uses support vector machines to obtain the region function type based on spatial positioning data and structural topology parameters. The region function type refers to the scene function category divided based on spatial positioning and topology structure. Then, the material risk level is obtained based on the region function type and optical feature parameters. The material risk level refers to the quantitative level of evaluating the flammability and hazard of object materials after burning. Finally, the spatial attribute category is obtained by considering the region function type, material risk level, and structural topology parameters. The spatial attribute category refers to a composite semantic label that integrates region function type, material risk level, and topology relationship. For example, the spatial positioning data (such as three-dimensional coordinates, area, and contour range) and structural topology parameters (such as object adjacency relationship and connectivity index) of the target region in the scene are first acquired through sensor collection and image analysis. Based on the above data, the support vector machine (SVM) classifier is used to determine the region function type, determining whether it belongs to the kitchen, electrical equipment room, warehouse area, or ordinary office area. Based on the determination of the region function type, the color mean, texture roughness, edge density, and other optical feature parameters extracted from the spectral image are combined to classify the fire risk level of the object material in the region through the decision tree algorithm, identifying the flammable material area, medium-risk material area, or low-risk material area. Finally, the spatial risk is corrected by integrating the region function type, material risk level, and spatial adjacency relationship (such as the distance to high-risk heat sources) in the structural topology parameters, thereby determining the final spatial attribute category of the region. Then, the color feature vector, dynamic texture entropy matrix, and thermodynamic gradient tensor are analyzed for correlation with the spatial attribute category, obtaining the corresponding modal correlation coefficients. The modal correlation coefficient is a statistical indicator that measures the degree of association between features and spatial attribute categories. The higher the coefficient, the more important the feature is to the risk assessment of the current scene.and a corresponding correlation weight index is obtained according to the modal correlation coefficient, wherein the correlation weight index refers to a weight factor dynamically assigned to each feature according to the modal correlation coefficient, and this step solves the fundamental defect of the traditional multi-modal fusion algorithm that "fixed weight cannot adapt to scene changes" through a dynamic weight distribution mechanism of scene perception, thereby significantly improving the accuracy and adaptability of fire warning, for example, if the correlation coefficient between the color feature vector and the spatial attribute category is , the correlation coefficient between the dynamic texture entropy matrix and the spatial attribute category is , and the correlation coefficient between the thermodynamic gradient tensor and the spatial attribute category is , then the modal correlation coefficients of the color feature vector, the dynamic texture entropy matrix, and the thermodynamic gradient tensor are , , , the color feature vector, the dynamic texture entropy matrix, and the thermodynamic gradient tensor have different data structures and dimensions, and direct splicing in the traditional method will cause feature expression confusion, therefore, the multi-modal data is normalized in the scheme to obtain a normalized feature dataset, for example: the color feature vector is subjected to Min-Max normalization to preserve the relative relationship of spectral features, the texture matrix is subjected to Z-score standardization to highlight abnormal texture changes, and the thermodynamic gradient tensor is subjected to norm normalization to maintain the physical meaning of spatial heat distribution, and finally the dynamic risk level is obtained by weighted summation combined with the weight index.

[0091] In one embodiment, the step S4 of obtaining the initial image processing parameters according to the computing performance data, the running performance data, and the dynamic risk level comprises:

[0092] S41, obtaining the core frequency, the reference frequency, the architecture coefficient, the utilization margin, the remaining memory capacity, and the memory fragmentation rate according to the computing performance data, and obtaining a resource comprehensive index according to the core frequency, the reference frequency, the architecture coefficient, and the utilization margin;

[0093] S42, obtaining a memory health degree according to the remaining memory capacity and the memory fragmentation rate;

[0094] S43, obtaining the real-time running temperature, the safety threshold temperature, and the maximum allowable temperature according to the running performance data, and obtaining a computing health degree according to the real-time running temperature, the safety threshold temperature, and the maximum allowable temperature;

[0095] S44, obtaining a resolution parameter according to the resource comprehensive index and the computing health degree;

[0096] S45, obtaining a risk-resource mapping table, and obtaining a computing demand benchmark and a memory demand benchmark according to the risk-resource mapping table and the dynamic risk level;

[0097] S46, obtain model architecture parameters according to the memory requirement benchmark and the memory health degree, and obtain calculation precision parameters according to the resource comprehensive index and the calculation requirement benchmark.

[0098] As described in steps S41-S46 above, in the present application, by obtaining core frequency, architecture coefficient and utilization margin, the three independent dimension parameters are fused into a unified index, breaking through the traditional single evaluation method based on frequency or utilization only, wherein the core frequency refers to the clock frequency of the processor core (such as CPU, GPU or NPU) in the embedded device, the architecture coefficient refers to a quantitative index representing the advancement of the processor architecture, and the utilization margin refers to the difference between the current utilization rate and the maximum utilization rate (usually 100%) of the processor core. Then, the resource comprehensive index is calculated by the formula "resource comprehensive index = 0.4*(core frequency / benchmark frequency) + 0.3*architecture coefficient + 0.7*utilization margin", wherein the resource comprehensive index refers to a quantitative value for evaluating the comprehensive computing power. Since the traditional method only focuses on the remaining memory amount, the present scheme considers the memory fragmentation rate at the same time, avoiding the problem of "having memory but unable to allocate" caused by memory fragmentation, and is particularly suitable for scenarios where the embedded system is short of memory. The memory health degree is calculated by the formula "memory health degree = 0.6*remaining memory capacity + 0.4*(1-memory fragmentation rate)", wherein the remaining memory capacity refers to the size of the physical memory space in the embedded device that is not occupied, the memory fragmentation rate refers to the proportion of non-continuous free blocks in the total free memory in the memory, and the memory health degree refers to a comprehensive index for measuring the availability of memory resources. Then, the temperature is included in the resource evaluation system, and the calculation load is actively reduced when approaching the safety threshold, avoiding the traditional passive mode of "overheating and frequency reduction→performance drop". First, the real-time running temperature is mapped to the [0, 1] interval by the formula "temperature pressure value = (real-time running temperature-safety threshold temperature) / (maximum allowed temperature-safety threshold temperature)", obtaining the temperature pressure value, which reflects the degree to which the current temperature approaches the dangerous area. Then, an exponential decay model is adopted to calculate the health degree: , calculating the calculation health degree, wherein the calculation health degree refers to a hardware running state evaluation index based on temperature. This method of converting hardware temperature physical constraints into a calculable resource index through a health degree function realizes the closed-loop optimization of "hardware state→algorithm parameter". The resolution parameter is calculated, wherein the resolution parameter refers to the input image resolution processed by the image recognition algorithm, and then the corresponding calculation requirement benchmark and memory requirement benchmark are obtained by table lookup according to the dynamic risk level and the risk-resource mapping table, wherein the risk-resource mapping table refers to a mapping relationship table of predefined dynamic risk levels (such as low, medium, and high risk) and hardware resource requirements, the calculation requirement benchmark refers to the minimum computing power requirement determined according to the risk level, and the memory requirement benchmark refers to the minimum memory requirement determined according to the risk level, so that the algorithm configuration is more suitable for the actual risk scene, and then the model architecture parameter is obtained according to the memory requirement benchmark and the memory health degree, wherein the model architecture parameter refers to a parameter representing the complexity of the image recognition model structure, such as the number of convolution layers, the number of channels, and the attention module configuration. For example, if the memory health degree is greater than or equal to 0.5: when the memory requirement benchmark is less than or equal to 512 MB, a lightweight or medium-weight model (such as EfficientNetB0) is selected to balance accuracy and performance; when the memory requirement benchmark is between 512 MB and 1024 MB, a medium-weight model (such as ResNet50y) is selected to ensure detection accuracy for common risks; and when the memory requirement benchmark is greater than or equal to 1024 MB, a complex model (such as YOLOv5s) is selected. Through the above method, the selection of the model architecture parameter is no longer dependent on fixed configuration, but is dynamically adjusted according to the real-time memory state and risk demand, which significantly improves the environmental adaptability of the embedded image recognition system. Finally, the calculation accuracy parameter is calculated according to the formula "calculation accuracy parameter = f (memory health degree, memory requirement benchmark)", wherein the calculation accuracy parameter refers to the data accuracy (such as FP32, FP16, and INT8) used in the algorithm calculation process. Through this method, when resources are insufficient, the accuracy can be gradually reduced instead of directly switching to a low-precision mode, thereby reducing the detection performance loss caused by sudden accuracy reduction.

[0099] In one embodiment, the step S5 of obtaining a real-time matching value according to the calculation performance data, risk assessment feature data, and initial image processing parameter comprises:

[0100] S51, obtaining hardware floating-point computing power effective value, chip junction temperature, cycle operation number, memory bandwidth, and processing period according to the calculation performance data, and obtaining CPU real-time computing power according to the hardware floating-point computing power effective value, core frequency, cycle operation number, chip junction temperature, and safety threshold temperature;

[0101] S52, obtaining processing requirement data according to the risk assessment feature data, and obtaining processing time threshold and algorithm calculation amount according to the processing requirement data;

[0102] S53, obtaining theoretical processing time according to the algorithm calculation amount and the CPU real-time computing power, and obtaining algorithm matching degree according to the theoretical processing time and the processing time threshold; ​

[0103] S54, obtaining memory available space according to the remaining memory capacity, memory fragmentation rate, memory bandwidth and processing period;

[0104] S55, obtaining memory requirement space and requirement bandwidth according to the resolution parameter, model architecture parameter and calculation precision parameter, and obtaining memory matching degree according to the memory requirement space, memory available space, memory bandwidth and requirement bandwidth;

[0105] S56, obtaining real-time matching value according to the algorithm matching degree and the memory matching degree.

[0106] As described in steps S51-S56 above, the present application obtains the CPU theoretical computing power by the product of the hardware floating-point computing power effective value, the core frequency, and the cycle operand, wherein the hardware floating-point computing power effective value refers to the stable floating-point operation capability that the hardware (such as CPU / GPU) can output in actual operation, and the cycle operand refers to the number of floating-point operations that the CPU can execute per clock cycle, and the CPU theoretical computing power is dynamically adjusted according to the difference between the chip junction temperature and the safety threshold temperature (such as reducing the frequency by a certain proportion when the junction temperature exceeds the safety threshold), to obtain the CPU real-time computing power, wherein the chip junction temperature refers to the actual temperature of the semiconductor junction inside the chip, and the CPU real-time computing power refers to the actual computing capability of the CPU at the current time. By this way of obtaining the CPU real-time computing power, the traditional static evaluation mode relying on the nominal computing power can be broken through, the real computing power of the hardware is quantified in real time through the frequency utilization rate and the temperature frequency reduction coefficient, the computing power fluctuation problem caused by the heat dissipation limitation of the embedded device is solved, then the algorithm complexity is derived from the risk assessment feature data (such as color feature vector, thermodynamic gradient tensor, and dynamic texture entropy value matrix), to obtain the processing requirement data, wherein the processing requirement data refers to the image processing task requirements derived from the risk assessment feature data, including indicators such as calculation amount, timeliness, and memory occupation, then the processing time threshold and the algorithm calculation amount are obtained according to the processing requirement data, wherein the processing time threshold refers to the maximum time consumption allowed by the system for the image processing task, and the algorithm calculation amount refers to the number of floating-point operations required to execute a specific image processing algorithm, and the theoretical processing time is obtained according to the ratio of the algorithm calculation amount to the CPU real-time computing power, then the algorithm matching degree is calculated according to the formula "algorithm matching degree = 1-max (theoretical processing time-processing time threshold, 0) / processing time threshold", wherein the algorithm matching degree refers to a quantitative indicator for measuring the matching degree of algorithm calculation requirements and CPU real-time computing power, and the algorithm matching degree result ranges between 0 and 1, and the higher the value, the more it meets the real-time requirements. By this processing method of associating multi-modal risk features with processing requirements (such as forcibly reducing the time threshold to 50ms in high-risk scenarios), the computing power allocation is more suitable for the actual threat level, thereby avoiding the problem of "low risk and high computing power waste, and high risk and low computing power missing detection", then the memory available space is calculated by the formula "memory available space = remaining memory capacity * (1-memory fragmentation rate) + memory bandwidth * processing period", wherein the memory available space refers to the actual available memory resources considering the memory fragmentation and bandwidth utilization rate, the memory bandwidth refers to the data transmission rate between the memory and the CPU, and the processing period refers to the time period required to complete one complete image processing task, then the memory requirement space is further obtained according to the resolution parameter, the model architecture parameter, and the calculation precision parameter, wherein the memory requirement space refers to the memory resources required to execute a specific image processing algorithm, and the memory matching degree is calculated according to the formula "memory matching degree = (memory available space / memory requirement space) * min (1, memory bandwidth / requirement bandwidth)", whereinThe demand bandwidth refers to the data transmission rate requirement between the algorithm runtime memory and the CPU, and the memory matching degree refers to a quantitative index for measuring the matching degree of the algorithm memory requirement and the hardware memory resource. By including the fragmentation rate (spatial continuity) and the bandwidth (transmission efficiency) into the evaluation system, the problem of "processing delay caused by insufficient bandwidth although the capacity is sufficient" in the embedded system can be solved, thereby breaking the traditional "capacity priority" memory management mode. Finally, the real-time matching value is obtained by fusing the algorithm matching degree and the memory matching degree according to the preset weight. The real-time matching value can be used to dynamically evaluate the adaptation degree of the hardware resources of the embedded image recognition system and the algorithm processing requirement, thereby providing a quantitative basis for system optimization, task scheduling or parameter adjustment.

[0107] In one embodiment, the step S6 of adjusting the initial image processing parameter according to the real-time matching value and the dynamic risk level comprises:

[0108] S61, obtaining an adjustment intensity level according to the real-time matching value and the dynamic risk level, and obtaining an adjustment constraint condition according to the dynamic risk level;

[0109] S62, obtaining a parameter adjustment data set according to the adjustment intensity level and the adjustment constraint condition, wherein the parameter adjustment data set comprises a plurality of resolution adjustment data, model architecture adjustment data and calculation precision adjustment data;

[0110] S63, arranging and combining a plurality of the resolution adjustment data, model architecture adjustment data and calculation precision adjustment data to obtain a plurality of parameter adjustment combination sets;

[0111] S64, screening a plurality of parameter adjustment combination sets according to the real-time matching value and the adjustment constraint condition to obtain an optimal parameter adjustment combination, and returning the optimal parameter adjustment combination to the step of obtaining the real-time matching value according to the calculation performance data, the risk evaluation feature data and the initial image processing parameter.

[0112] As described in steps S61-S64 above, the application breaks through the limitation of single factor determining adjustment strategy by constructing a two-dimensional mapping table of real-time matching value and dynamic risk level, realizes more fine adjustment intensity division, for example: dividing the real-time matching value into low (0-0.3), medium (0.3-0.7), and high (0.7-1) three intervals, dividing the dynamic risk level into 1-5 levels, determining the corresponding adjustment intensity level through the cross-index mapping table, wherein the adjustment intensity level refers to the parameter adjustment intensity classification determined according to the real-time matching value and the dynamic risk level, for example, when the real-time matching value is 0.2 and the dynamic risk level is 4, the highest adjustment intensity level is obtained by looking up the table, and at the same time, the adjustment constraint condition is obtained according to the dynamic risk level, wherein the adjustment constraint condition refers to the hard limit condition determined by the dynamic risk level, which is used to ensure that the parameter adjustment is within a safe and feasible range, for example: in a low-risk scenario, a lower resolution (such as not less than 320x240), a lower model accuracy (such as INT8 quantization), and a longer delay (such as not more than 500ms) are allowed; in a high-risk scenario, a higher resolution (such as not less than 1080P), a higher model accuracy (such as FP16), and a shorter delay (such as not more than 100ms) are required, in this way, the necessary image processing quality and system response speed can be maintained in a high-risk scenario, avoiding problems such as missed detection caused by excessive adjustment, then the parameter adjustment dataset is obtained according to the adjustment intensity level and the adjustment constraint condition, wherein the parameter adjustment dataset refers to a basic data set containing various adjustable parameters, including multiple resolution adjustment data, model architecture adjustment data, and calculation precision adjustment data, this acquisition method avoids blind enumeration of all possible parameters, reducing invalid calculations, then a permutation and combination algorithm is used to combine the resolution adjustment data, model architecture adjustment data, and calculation precision adjustment data, for example, if there are 3 resolution values, 2 model architectures, and 2 calculation precision values, then 3x2x2=12 parameter adjustment combination sets can be generated, wherein the parameter adjustment combination set refers to a plurality of candidate adjustment schemes generated by permutation and combination of resolution, model architecture, calculation precision, etc. parameters in the parameter adjustment dataset, each combination set contains a complete set of resolution, model architecture, and calculation precision parameters, finally, for each parameter adjustment combination set, the real-time matching value and the adjustment constraint condition are combined, and a multi-objective optimization algorithm (such as non-dominated sorting genetic algorithm NSGA-II) is used for evaluation, the combinations that meet the constraints are sorted, and the combination with the highest real-time matching value is selected as the optimal parameter adjustment combination, and the optimal parameter adjustment combination is returned to the step of obtaining the real-time matching value as the initial image processing parameter, used to recalculate the real-time matching value and verify the adjustment effect, through this closed-loop adjustment method, the system can be continuously optimized, which is conducive to adapting to the dynamic changes of the environment and system state.

[0113] The application also provides an optimization system for an embedded image recognition algorithm, comprising:

[0114] A first acquisition module is configured to acquire multi-modal monitoring data monitored by a sensor, wherein the multi-modal monitoring data comprises scene feature data, multi-modal environment data, and hardware state data;

[0115] A second acquisition module is configured to acquire risk assessment feature data according to the multi-modal environment data;

[0116] A third acquisition module is configured to acquire a dynamic risk level according to the risk assessment feature data and the scene feature data;

[0117] A fourth acquisition module is configured to acquire computing performance data and running performance data according to the hardware state data, and to acquire initial image processing parameters according to the computing performance data, the running performance data, and the dynamic risk level;

[0118] A fifth acquisition module is configured to acquire a real-time matching value according to the computing performance data, the risk assessment feature data, and the initial image processing parameters;

[0119] A judgment module is configured to judge whether the real-time matching value is greater than a preset threshold value;

[0120] If the real-time matching value is greater than the preset threshold value, it is determined that the calculation delay meets the real-time requirement;

[0121] If the real-time matching value is not greater than the preset threshold value, it is determined that the calculation delay does not meet the real-time requirement, and the initial image processing parameters are adjusted according to the real-time matching value and the dynamic risk level until the real-time matching value is greater than the preset threshold value.

[0122] In one embodiment, the third acquisition module comprises:

[0123] A first acquisition unit is configured to acquire spatial positioning data, optical feature parameters, and structure topology parameters according to the scene feature data, and to acquire a region function type according to the spatial positioning data and the structure topology parameters;

[0124] A second acquisition unit is configured to acquire a material risk level according to the region function type and the optical feature parameters, and to acquire a spatial attribute category according to the region function type, the material risk level, and the structure topology parameters;

[0125] A third acquisition unit is configured to perform correlation analysis on the color feature vector, the dynamic texture entropy value matrix, and the thermodynamic gradient tensor respectively with the spatial attribute category to obtain corresponding modal correlation coefficients, and to acquire corresponding correlation weight indexes according to the modal correlation coefficients;

[0126] The fourth acquisition unit is used to acquire a normalized feature data set according to the color feature vector, the dynamic texture entropy matrix and the thermodynamic gradient tensor, and to acquire a dynamic risk level according to the normalized feature data set and the association weight index.

[0127] The present invention also provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above-mentioned optimization method for embedded image recognition algorithm when executing the computer program.

[0128] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-mentioned optimization method for an embedded image recognition algorithm.

[0129] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided in this application and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct RAM bus dynamic RAM (DRDRAM), and RAM bus dynamic RAM (RDRAM).

[0130] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, apparatus, article, or method comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, apparatus, article, or method. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, apparatus, article, or method comprising the element.

[0131] The above merely provides the preferred embodiments of the present application, and is not intended to limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made according to the content of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. An optimization method for embedded image recognition algorithm, characterized in that: include: Acquiring multimodal monitoring data monitored by sensors, wherein the multimodal monitoring data includes scene feature data, multimodal environment data, and hardware status data; Obtaining risk assessment feature data based on the multimodal environmental data, wherein obtaining risk assessment feature data based on the multimodal environmental data includes: obtaining spectral image data, thermal radiation image data, and gaseous molecule data based on the multimodal environmental data, and obtaining a spectral alignment image and a thermal radiation alignment image based on the spectral image data and the thermal radiation image data; obtaining a global chromaticity variance based on the spectral alignment image, and obtaining a spectral feature vector based on the global chromaticity variance through principal component analysis; obtaining timestamp feature points based on the spectral alignment image and the thermal radiation alignment image, and obtaining time-aligned gas data based on the timestamp feature points and the gaseous molecule data; obtaining a CO concentration sequence and a concentration gradient based on the time-aligned gas data, and obtaining a color feature vector based on the CO concentration sequence, the concentration gradient, and the spectral feature vector; obtaining a high-temperature centroid coordinate and a temperature gradient matrix based on the thermal radiation alignment image, and obtaining a spatiotemporal adjacency matrix based on the high-temperature centroid coordinate; obtaining a thermodynamic gradient tensor based on the spatiotemporal adjacency matrix and the temperature gradient matrix, and obtaining a dynamic texture entropy matrix based on the spatiotemporal adjacency matrix; Obtaining a dynamic risk level based on the risk assessment feature data and the scenario feature data; Obtaining computing performance data and operating performance data according to the hardware status data, and obtaining initial image processing parameters according to the computing performance data, operating performance data, and dynamic risk level; obtaining a real-time matching value based on the computing performance data, the risk assessment feature data, and the initial image processing parameters; Determine whether the real-time matching value is greater than a preset threshold; If the real-time matching value is greater than a preset threshold, it is determined that the calculation delay meets the real-time requirement; If the real-time matching value is not greater than the preset threshold, it is determined that the calculation delay does not meet the real-time requirement, and the initial image processing parameters are adjusted according to the real-time matching value and the dynamic risk level until the real-time matching value is greater than the preset threshold.

2. The optimization method for embedded image recognition algorithm according to claim 1, characterized in that: The step of obtaining a dynamic risk level according to the risk assessment feature data and the scenario feature data includes: Acquire spatial positioning data, optical characteristic parameters and structural topology parameters according to the scene characteristic data, and acquire regional functional type according to the spatial positioning data and the structural topology parameters; Obtaining a material risk level according to the regional function type and the optical characteristic parameters, and obtaining a spatial attribute category according to the regional function type, the material risk level, and the structural topology parameters; Performing correlation analysis on the color feature vector, dynamic texture entropy matrix, and thermodynamic gradient tensor with the spatial attribute category to obtain corresponding modal correlation coefficients, and obtaining corresponding association weight indexes according to the modal correlation coefficients; A normalized feature data set is obtained according to the color feature vector, the dynamic texture entropy matrix and the thermodynamic gradient tensor, and a dynamic risk level is obtained according to the normalized feature data set and the associated weight index.

3. The optimization method for embedded image recognition algorithm according to claim 1, characterized in that: The step of obtaining initial image processing parameters according to the computing performance data, the operating performance data and the dynamic risk level includes: Obtaining a core frequency, a reference frequency, an architecture coefficient, a utilization margin, a remaining memory capacity, and a memory fragmentation rate according to the computing performance data, and obtaining a resource comprehensive index according to the core frequency, the reference frequency, the architecture coefficient, and the utilization margin; Obtaining memory health according to the remaining memory capacity and the memory fragmentation rate; Obtaining a real-time operating temperature, a safety threshold temperature, and a maximum allowable temperature according to the operating performance data, and calculating a health level according to the real-time operating temperature, the safety threshold temperature, and the maximum allowable temperature; Obtaining a resolution parameter according to the resource comprehensive index and the computing health; Obtaining a risk-resource mapping table, and obtaining a computing requirement benchmark and a memory requirement benchmark based on the risk-resource mapping table and the dynamic risk level; Model architecture parameters are obtained according to the memory requirement benchmark and the memory health, and computing accuracy parameters are obtained according to the resource comprehensive index and the computing requirement benchmark.

4. The optimization method for embedded image recognition algorithm according to claim 3, characterized in that: The step of obtaining a real-time matching value based on the computing performance data, the risk assessment feature data, and the initial image processing parameters includes: Obtaining a hardware floating-point computing power effective value, chip junction temperature, cycle operation number, memory bandwidth, and processing cycle according to the computing performance data, and obtaining a CPU real-time computing power according to the hardware floating-point computing power effective value, core frequency, cycle operation number, chip junction temperature, and safety threshold temperature; Acquiring processing requirement data based on the risk assessment feature data, and acquiring a processing time threshold and an algorithm calculation amount based on the processing requirement data; Obtaining a theoretical processing time based on the algorithm calculation amount and the CPU real-time computing power, and obtaining an algorithm matching degree based on the theoretical processing time and the processing time threshold; Obtaining memory available space based on the remaining memory capacity, memory fragmentation rate, memory bandwidth, and processing cycle; Obtaining a memory requirement space and a required bandwidth according to the resolution parameter, the model architecture parameter, and the calculation accuracy parameter, and obtaining a memory matching degree according to the memory requirement space, the available memory space, the memory bandwidth, and the required bandwidth; A real-time matching value is obtained according to the algorithm matching degree and the memory matching degree.

5. The optimization method for embedded image recognition algorithm according to claim 1, characterized in that: The step of adjusting the initial image processing parameters according to the real-time matching value and the dynamic risk level includes: Obtaining an adjustment intensity level according to the real-time matching value and the dynamic risk level, and obtaining an adjustment constraint condition according to the dynamic risk level; Acquire a parameter adjustment data set according to the adjustment intensity level and the adjustment constraint condition, wherein the parameter adjustment data set includes a plurality of resolution adjustment data, model architecture adjustment data, and calculation accuracy adjustment data; Arrange and combine the plurality of resolution adjustment data, model architecture adjustment data, and calculation accuracy adjustment data to obtain a plurality of parameter adjustment combination sets; A plurality of parameter adjustment combination sets are screened according to the real-time matching value and the adjustment constraint condition to obtain an optimal parameter adjustment combination, and the optimal parameter adjustment combination is returned as the initial image processing parameter to the step of obtaining the real-time matching value according to the computing performance data, the risk assessment feature data and the initial image processing parameter.

6. An optimization system for embedded image recognition algorithms, characterized in that: include: A first acquisition module is used to acquire multimodal monitoring data monitored by sensors, wherein the multimodal monitoring data includes scene feature data, multimodal environment data and hardware status data; a second acquisition module, configured to acquire risk assessment feature data based on the multimodal environmental data, wherein acquiring risk assessment feature data based on the multimodal environmental data comprises: acquiring spectral image data, thermal radiation image data, and gaseous molecule data based on the multimodal environmental data, and acquiring a spectral alignment image and a thermal radiation alignment image based on the spectral image data and the thermal radiation image data; acquiring global chromaticity variance based on the spectral alignment image, and acquiring a spectral feature vector based on the global chromaticity variance through principal component analysis; acquiring timestamp feature points based on the spectral alignment image and the thermal radiation alignment image, and acquiring time-aligned gas data based on the timestamp feature points and the gaseous molecule data; acquiring a CO concentration sequence and a concentration gradient based on the time-aligned gas data, and acquiring a color feature vector based on the CO concentration sequence, the concentration gradient, and the spectral feature vector; acquiring high-temperature centroid coordinates and a temperature gradient matrix based on the thermal radiation alignment image, and acquiring a spatiotemporal adjacency matrix based on the high-temperature centroid coordinates; acquiring a thermodynamic gradient tensor based on the spatiotemporal adjacency matrix and the temperature gradient matrix, and acquiring a dynamic texture entropy matrix based on the spatiotemporal adjacency matrix; a third acquisition module, configured to acquire a dynamic risk level based on the risk assessment feature data and the scene feature data; specifically, the third acquisition module includes: a first acquisition unit, configured to acquire spatial positioning data, optical feature parameters, and structural topology parameters based on the scene feature data, and to acquire a regional functional type based on the spatial positioning data and the structural topology parameters; a second acquisition unit, configured to acquire a material risk level based on the regional functional type and the optical feature parameters, and to acquire a spatial attribute category based on the regional functional type, the material risk level, and the structural topology parameters; a third acquisition unit, configured to perform correlation analysis on the color feature vector, the dynamic texture entropy matrix, and the thermodynamic gradient tensor with the spatial attribute category, respectively, to obtain a corresponding modal correlation coefficient, and to obtain a corresponding association weight index based on the modal correlation coefficient; a fourth acquisition unit, configured to acquire a normalized feature data set based on the color feature vector, the dynamic texture entropy matrix, and the thermodynamic gradient tensor, and to acquire a dynamic risk level based on the normalized feature data set and the association weight index; a fourth acquisition module, configured to acquire computing performance data and operating performance data according to the hardware status data, and to acquire initial image processing parameters according to the computing performance data, operating performance data, and dynamic risk level; a fifth acquisition module, configured to acquire a real-time matching value based on the computing performance data, the risk assessment feature data, and the initial image processing parameters; A judgment module, used to judge whether the real-time matching value is greater than a preset threshold; If the real-time matching value is greater than a preset threshold, it is determined that the calculation delay meets the real-time requirement; If the real-time matching value is not greater than the preset threshold, it is determined that the calculation delay does not meet the real-time requirement, and the initial image processing parameters are adjusted according to the real-time matching value and the dynamic risk level until the real-time matching value is greater than the preset threshold.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 5 are implemented.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

Images