Airborne cloud particle image quality intelligent evaluation and screening method and system

Abstract

This invention relates to the field of airborne cloud particle image processing technology, and provides a method and system for intelligent assessment and screening of airborne cloud particle image quality. The method extracts comprehensive quality features through multi-scale spatial and frequency domain joint analysis, employs a multi-task learning architecture to calculate image quality scores and defect types, determines a quality screening threshold based on score distribution characteristics and environmental parameters, retains valid images above the threshold and removes invalid images, and updates the learning architecture parameters based on the distribution of quality features of valid images and the clustering results of defect features of invalid images. This invention improves the accuracy and adaptability of cloud particle image screening.

Description

Technical Field

[0001] This invention relates to the field of airborne cloud particle image processing technology, and in particular to an intelligent evaluation and screening method and system for airborne cloud particle image quality. Background Technology

[0002] Airborne cloud particle detection is an important tool in atmospheric science and meteorological research. It involves acquiring image data of particles in clouds using optical detection equipment carried on an aircraft platform, which is then used to analyze cloud microphysical properties, water vapor distribution, and climate change. However, during actual flight, due to complex flight environments, equipment vibration, and variations in lighting, the quality of the acquired cloud particle images varies considerably. Many images are low-quality, including blurry, underexposed, overexposed, and noise-affected images, severely impacting subsequent scientific analysis and data processing.

[0003] Traditional airborne cloud particle image quality assessment relies primarily on manual inspection or basic image processing algorithms for screening. Researchers typically need to manually review a large number of images and subjectively judge their quality, resulting in a huge workload and inconsistent evaluation standards. Some automated methods use only simple sharpness indicators or brightness statistics for evaluation, which cannot comprehensively reflect the complex quality characteristics of cloud particle images. Existing technologies have the following shortcomings in cloud particle image quality assessment and screening: Most existing methods are based on feature extraction at a single scale or in a single domain, making it difficult to comprehensively capture the various quality defects in cloud particle images. Cloud particle image quality problems manifest in multiple aspects, such as edge sharpness in the spatial domain and spectral distribution in the frequency domain; a single feature is insufficient to accurately describe the overall image quality status. Existing quality assessment models typically use static thresholds for screening, failing to consider the dynamic changes in meteorological environments at different flight stages. In actual detection processes, the distribution pattern of image quality constantly changes with flight altitude, cloud density, and atmospheric conditions, making fixed thresholds unsuitable for this dynamic environment. Existing technologies lack the ability to accurately identify and classify image quality defect types, only providing vague quality level classifications. In practical applications, identifying specific types of quality defects (such as motion blur, insufficient lighting, excessive particle density, etc.) is of great significance for improving the parameters of the acquisition equipment and post-processing strategies. Summary of the Invention

[0004] This invention provides a method and system for intelligent evaluation and screening of airborne cloud particle image quality, which can solve the problems in the prior art.

[0005] A first aspect of this invention provides an intelligent evaluation and screening method for airborne cloud particle image quality, comprising:

[0006] Multi-scale spatial and frequency domain joint analysis was performed on each image in the raw cloud particle image sequence acquired by the airborne cloud particle detection equipment to extract the comprehensive quality feature vector;

[0007] Based on the comprehensive quality feature vector, a multi-task learning architecture is used to simultaneously calculate the quality score and key quality defect types for each image;

[0008] Based on the statistical distribution characteristics of the quality score, and combined with the meteorological environmental parameters and data acquisition rate of the current flight phase, a quality screening threshold matching the current acquisition status is calculated;

[0009] Images in the original cloud particle image sequence with a quality score higher than the quality screening threshold are marked as valid images and retained, while images with a quality score lower than the quality screening threshold are marked as invalid images and removed, resulting in a filtered image set.

[0010] The quality feature distribution pattern of valid images and the defect feature clustering result of invalid images in the filtered image set are determined. Based on the quality feature distribution pattern and the defect feature clustering result, the feature weight coefficients and classification decision boundary in the multi-task learning architecture are updated.

[0011] Multi-scale spatial and frequency domain joint analysis was performed on each image in the raw cloud particle image sequence acquired by the airborne cloud particle detection equipment to extract a comprehensive quality feature vector, including:

[0012] Each image in the original cloud particle image sequence is subjected to multi-scale spatial decomposition to obtain spatial domain quality feature sub-vectors;

[0013] Perform a frequency domain transformation on each image in the original cloud particle image sequence to obtain a frequency domain quality feature subvector;

[0014] Based on the high-frequency gradient components in the spatial domain quality feature vector and the high-frequency energy distribution characteristics in the frequency domain quality feature vector, particle edge information and noise interference components in the image are distinguished, and a noise intensity assessment value characterizing the noise level is calculated.

[0015] Based on the edge continuity metric in the spatial domain quality feature sub-vector, and combined with the noise intensity evaluation value, the edge detection results are weighted for reliability. By constructing the connectivity topology of edge pixels and the consistency constraint of edge intensity, the edge quality evaluation value characterizing the integrity of particle edges is calculated.

[0016] The spatial domain quality feature vector, the frequency domain quality feature vector, the noise intensity assessment value, and the edge quality assessment value are weighted and combined to obtain the comprehensive quality feature vector.

[0017] Based on the statistical distribution characteristics of the quality score, and combined with the meteorological environmental parameters and data acquisition rate of the current flight phase, a quality screening threshold matching the current acquisition status is calculated, including:

[0018] Statistical analysis was performed on the quality scores of all images in the original cloud particle image sequence to obtain the statistical characteristics of the quality scores;

[0019] Obtain meteorological environmental parameters and data acquisition rate for the current flight phase. Based on the correlation between the meteorological environmental parameters and the statistical characteristics of quality scores in historical flight phases, determine the quantitative relationship of the influence of meteorological environmental conditions on the quality distribution of cloud particle images, and calculate the expected quality offset under the current meteorological environmental conditions.

[0020] Based on the data acquisition rate, the available capacity of onboard storage resources, and the minimum requirement for the number of valid images, the target data retention rate under the current acquisition state is calculated by combining storage capacity constraints and data requirement constraints.

[0021] Based on the quantile distribution curve in the quality score statistical features, the quantile distribution curve is corrected by combining the expected quality offset;

[0022] Find the quantiles on the corrected quantile distribution curve that correspond to the target data retention rate, and determine the quality score of the quantiles as the quality screening threshold.

[0023] Obtain meteorological environmental parameters and data acquisition rate for the current flight phase. Based on the correlation between the meteorological environmental parameters and the statistical characteristics of quality scores in historical flight phases, determine the quantitative relationship of the impact of meteorological environmental conditions on cloud particle image quality distribution. Calculate the expected quality offset under the current meteorological environmental conditions, including:

[0024] Meteorological environmental parameters and data acquisition rates for the current flight phase are obtained from airborne meteorological sensors and flight control systems, and meteorological environmental parameter records and corresponding quality score statistical feature records for multiple historical flight phases are extracted from the historical flight database.

[0025] For each historical flight phase, a functional mapping relationship is determined between the meteorological environmental parameter records and the quality score statistical feature records. A quantitative expression for the influence of meteorological environmental conditions on the quality score distribution is then determined by fitting the functional mapping relationship.

[0026] Based on the meteorological environmental parameters of the current flight phase and the quantitative relationship expression, the expected quality score distribution characteristics under the current meteorological environmental conditions are determined.

[0027] The deviation between the expected quality score distribution characteristics and the actual quality score statistical characteristics of the current flight phase is determined as the expected quality offset.

[0028] The expected mass offset is determined based on the ratio between the data acquisition rate of the current flight phase and the data acquisition rate of the historical flight phase.

[0029] Images in the original cloud particle image sequence with quality scores higher than the quality screening threshold are marked as valid images and retained, while images with quality scores lower than the quality screening threshold are marked as invalid images and removed, resulting in a filtered image set, including:

[0030] The quality score of each image in the original cloud particle image sequence is compared with the quality screening threshold, and a screening label for each image is determined based on the comparison result.

[0031] Identify boundary images where the absolute value of the difference between the quality score and the quality screening threshold is less than a preset boundary range, and extract key quality defect type information from the boundary images.

[0032] Determine whether the critical quality defect type belongs to the set of tolerable defect types, and adjust the filtering markers of the boundary image based on the determination result;

[0033] The time interval between adjacent images that are screened as valid images is detected. When the time interval exceeds a continuity threshold, the candidate compensation image with the highest quality score is searched among the invalid images located between the adjacent images in the original cloud particle image sequence.

[0034] Determine whether the quality score of the candidate compensated image meets the downgrade retention condition. If it does, correct the screening mark of the candidate compensated image to a valid image mark. Organize all images with the screening mark as valid image marks in the order of acquisition time to obtain the filtered image set.

[0035] Determine the quality feature distribution pattern of valid images and the defect feature clustering result of invalid images in the filtered image set. Based on the quality feature distribution pattern and the defect feature clustering result, update the feature weight coefficients and classification decision boundary in the multi-task learning architecture, including:

[0036] Based on the comprehensive quality feature vector of all valid images in the filtered image set, calculate the quality feature distribution pattern of the valid images;

[0037] The defect feature clustering results of the invalid images are determined based on the key quality defect types and comprehensive quality feature vectors corresponding to all invalid images in the original cloud particle image sequence;

[0038] For each feature dimension, the minimum distance between the mean in the quality feature distribution pattern and the feature center of each defect category in the defect feature clustering result is calculated, and the feature discrimination index is calculated based on the minimum distance and the quality feature distribution pattern.

[0039] The feature weight coefficients of the corresponding feature dimensions in the multi-task learning architecture are adjusted according to the value of the feature discrimination index, so that the feature weight coefficients are positively correlated with the feature discrimination index.

[0040] Based on the feature center location and feature dispersion of each defect category in the defect feature clustering results, the classification decision boundary corresponding to the defect category is adjusted in the defect recognition decision layer of the multi-task learning architecture so that the classification decision boundary matches the actual feature distribution range of the defect category.

[0041] The defect feature clustering results of the invalid images are determined based on the key quality defect types and comprehensive quality feature vectors corresponding to all invalid images in the original cloud particle image sequence, including:

[0042] Traverse all invalid images in the original cloud particle image sequence, extract the key quality defect types corresponding to each invalid image, count the frequency of occurrence of each key quality defect type, and determine the set of all defect types existing in the original cloud particle image sequence;

[0043] For each defect type in the set of all defect types, extract the comprehensive quality feature vectors corresponding to all invalid images of the original cloud particle image sequence whose key quality defect type is that defect type, and obtain the feature vector set corresponding to the defect type;

[0044] For each defect type, a feature center is calculated based on the comprehensive quality feature vectors in the feature vector set.

[0045] For each defect type, calculate the distance between each comprehensive quality feature vector in the feature vector set corresponding to the defect type and the feature center of the defect type, and determine the feature dispersion of the defect type based on the distance value;

[0046] The feature center and feature dispersion corresponding to each defect type in the set of all defect types are used as the defect feature clustering result.

[0047] A second aspect of the present invention provides an airborne cloud particle image quality intelligent assessment and screening system, comprising:

[0048] The first unit is used to perform multi-scale spatial domain and frequency domain joint analysis on each image in the original cloud particle image sequence acquired by the airborne cloud particle detection equipment, and extract the comprehensive quality feature vector;

[0049] The second unit is used to simultaneously calculate the quality score and key quality defect types of each image based on the comprehensive quality feature vector using a multi-task learning architecture;

[0050] The third unit is used to calculate a quality screening threshold that matches the current data acquisition status based on the statistical distribution characteristics of the quality score, combined with the meteorological environmental parameters and data acquisition rate of the current flight phase.

[0051] The fourth unit is used to mark images in the original cloud particle image sequence with a quality score higher than the quality screening threshold as valid images and retain them, and to mark images with a quality score lower than the quality screening threshold as invalid images and remove them, thereby obtaining a set of filtered images;

[0052] The fifth unit is used to determine the quality feature distribution pattern of valid images and the defect feature clustering result of invalid images in the filtered image set, and to update the feature weight coefficients and classification decision boundary in the multi-task learning architecture based on the quality feature distribution pattern and the defect feature clustering result.

[0053] A third aspect of the present invention provides an electronic device, comprising:

[0054] processor;

[0055] Memory used to store processor-executable instructions;

[0056] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0057] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0058] This invention extracts a comprehensive quality feature vector through joint analysis in the multi-scale spatial and frequency domains, overcoming the limitations of evaluating image quality using a single feature dimension. It comprehensively captures various quality defect features in cloud particle images, improving the accuracy and comprehensiveness of quality assessment. This invention employs a multi-task learning architecture to simultaneously calculate quality scores and key quality defect types, achieving unified processing of quantitative image quality assessment and quality defect type identification. Compared to traditional single-score methods, it provides richer quality information, facilitating subsequent analysis and processing. This invention dynamically calculates screening thresholds based on the statistical distribution characteristics of quality scores, combined with meteorological environmental parameters and data acquisition rates during the flight phase. This achieves adaptive matching between screening criteria and the real-time acquisition environment, avoiding the inadequacy of fixed-threshold screening methods in complex and changing environments, and improving the reliability of screening results. This invention achieves adaptive optimization and updating of the evaluation model by analyzing the quality feature distribution of effective images and the defect feature clustering results of invalid images in the filtered image set. This allows the model to continuously learn and adapt to new image quality features and defect patterns, improving the method's adaptability and robustness. This invention integrates image quality assessment, defect type identification, dynamic threshold screening, and model adaptive updating into a complete processing flow, which can significantly improve the quality and availability of airborne cloud particle image data, providing a high-quality data foundation for subsequent cloud particle characteristic analysis and atmospheric environment research. Attached Figure Description

[0059] Figure 1 This is a flowchart illustrating the airborne cloud particle image quality intelligent evaluation and screening method according to an embodiment of the present invention.

[0060] Figure 2 This is a schematic diagram of the process for determining the expected quality offset according to an embodiment of the present invention. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0063] Figure 1 This is a flowchart illustrating the intelligent assessment and screening method for airborne cloud particle image quality according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0064] Multi-scale spatial and frequency domain joint analysis was performed on each image in the raw cloud particle image sequence acquired by the airborne cloud particle detection equipment to extract the comprehensive quality feature vector;

[0065] Based on the comprehensive quality feature vector, a multi-task learning architecture is used to simultaneously calculate the quality score and key quality defect types for each image;

[0066] Based on the statistical distribution characteristics of the quality score, and combined with the meteorological environmental parameters and data acquisition rate of the current flight phase, a quality screening threshold matching the current acquisition status is calculated;

[0067] Images in the original cloud particle image sequence with a quality score higher than the quality screening threshold are marked as valid images and retained, while images with a quality score lower than the quality screening threshold are marked as invalid images and removed, resulting in a filtered image set.

[0068] The quality feature distribution pattern of valid images and the defect feature clustering result of invalid images in the filtered image set are determined. Based on the quality feature distribution pattern and the defect feature clustering result, the feature weight coefficients and classification decision boundary in the multi-task learning architecture are updated.

[0069] In one optional implementation, each image in the raw cloud particle image sequence acquired by the airborne cloud particle detection equipment undergoes multi-scale spatial and frequency domain joint analysis to extract a comprehensive quality feature vector, including:

[0070] Each image in the original cloud particle image sequence is subjected to multi-scale spatial decomposition to obtain spatial domain quality feature sub-vectors;

[0071] Perform a frequency domain transformation on each image in the original cloud particle image sequence to obtain a frequency domain quality feature subvector;

[0072] Based on the high-frequency gradient components in the spatial domain quality feature vector and the high-frequency energy distribution characteristics in the frequency domain quality feature vector, particle edge information and noise interference components in the image are distinguished, and a noise intensity assessment value characterizing the noise level is calculated.

[0073] Based on the edge continuity metric in the spatial domain quality feature sub-vector, and combined with the noise intensity evaluation value, the edge detection results are weighted for reliability. By constructing the connectivity topology of edge pixels and the consistency constraint of edge intensity, the edge quality evaluation value characterizing the integrity of particle edges is calculated.

[0074] The spatial domain quality feature vector, the frequency domain quality feature vector, the noise intensity assessment value, and the edge quality assessment value are weighted and combined to obtain the comprehensive quality feature vector.

[0075] First, each image in the original cloud particle image sequence undergoes multi-scale spatial decomposition to obtain spatial domain quality feature sub-vectors. Specifically, wavelet transform is used to decompose the images into low-frequency approximation components and high-frequency detail components. In practice, Haar wavelet or Daubechies wavelet is chosen for three-level decomposition, resulting in four sub-bands: low-frequency approximation coefficient LL, horizontal detail coefficient LH, vertical detail coefficient HL, and diagonal detail coefficient HH. The spatial domain quality feature sub-vectors are constructed by calculating the statistical properties (such as mean, variance, kurtosis, and skewness) of each sub-band. Simultaneously, edge information is detected using edge gradient operators in different directions (such as Sobel, Prewitt, or Canny), and the distribution characteristics of gradient magnitudes are calculated as components of the spatial domain feature sub-vectors.

[0076] Secondly, each image in the original cloud particle image sequence undergoes a frequency domain transformation to obtain a frequency domain quality feature sub-vector. Specifically, a two-dimensional discrete Fourier transform (2D-DFT) is used to transform the image from the spatial domain to the frequency domain. In the frequency domain, the spectrum is divided into multiple annular and sector regions, representing texture features of different frequency components and directions, respectively. By calculating the energy distribution, energy concentration, and directionality of each region, the frequency domain quality feature sub-vector is constructed. To improve computational efficiency, a two-dimensional discrete cosine transform (2D-DCT) can also be used as an alternative.

[0077] Next, based on the high-frequency gradient components in the spatial domain quality feature vector and the high-frequency energy distribution characteristics in the frequency domain quality feature vector, particle edge information and noise interference components in the image are distinguished, and a noise intensity assessment value representing the noise level is calculated. Specifically, in the spatial domain, true particle edges usually exhibit gradient features with consistent direction and continuous intensity, while noise appears as randomly distributed, isolated gradient points. Simultaneously, in the frequency domain, noise typically manifests as an anomalous energy enhancement in high-frequency regions. By establishing a mapping relationship between spatial domain gradient features and frequency domain energy features, the noise intensity assessment value N_score is calculated.

[0078] By comparing the spatial clustering of high-frequency gradient points with the energy proportion of high-frequency regions in the frequency domain, and combining this with directional consistency analysis, a noise intensity assessment value is calculated. The larger this value, the stronger the image noise and the lower the quality. In the actual calculation, a threshold is first set to distinguish between potential noise points and edge points. Then, the proportion of noise points in the entire image is statistically analyzed, and combined with the proportion of high-frequency energy in the frequency domain, the final noise intensity assessment value is obtained.

[0079] Furthermore, based on the edge continuity metric in the spatial domain quality feature sub-vector, and combined with the noise intensity evaluation value, the edge detection results are weighted for reliability. By constructing the connectivity topology of edge pixels and the consistency constraint of edge intensity, an edge quality evaluation value characterizing the integrity of particle edges is calculated. First, an improved Canny edge detector is used to obtain an initial edge map. Then, the local connectivity and orientation consistency of each edge pixel are calculated. For each detected edge line, its integrity index, including length, curvature change, and closure, is calculated. Finally, the edge quality is weighted and adjusted using the previously obtained noise intensity evaluation value N_score to obtain the edge quality evaluation value E_score. A higher edge quality evaluation value indicates a more complete and clearer particle edge contour, and better image quality.

[0080] Finally, the spatial domain quality feature vector, frequency domain quality feature vector, noise intensity assessment value, and edge quality assessment value are weighted and combined to obtain the comprehensive quality feature vector. Weight coefficients w1, w2, w3, and w4 are set for the four parts, and the final comprehensive quality feature vector Q is constructed by weighted summation. The weight coefficients can be determined using training datasets and machine learning methods, or they can be manually adjusted according to application requirements. In practical applications, different weight configuration schemes can be adopted for different types of cloud particles (such as raindrops and ice crystals) to adapt to the characteristics of different particles.

[0081] In practical applications, when airborne cloud particle detection equipment collects cloud particle images at high speeds, the quality of the acquired raw images varies due to factors such as vibration and changes in lighting conditions. The comprehensive quality feature vector extracted using this method can be used to assess and filter the image sequence, providing a reliable data foundation for subsequent cloud particle identification, classification, and measurement. For example, in a high-altitude cloud sampling process, this method was used to assess the quality of three thousand cloud particle images, selecting higher-quality images for subsequent analysis, effectively improving the accuracy of cloud particle attribute measurements.

[0082] The advantage of this method lies in combining complementary information from the spatial and frequency domains, improving feature representation through multi-scale analysis, effectively distinguishing real particle edges from noise interference, and providing a reliable basis for image quality assessment for airborne cloud particle detection.

[0083] In one optional implementation, based on the statistical distribution characteristics of the quality score, and in conjunction with the meteorological environmental parameters and data acquisition rate of the current flight phase, a quality screening threshold matching the current acquisition status is calculated, including:

[0084] Statistical analysis was performed on the quality scores of all images in the original cloud particle image sequence to obtain the statistical characteristics of the quality scores;

[0085] Obtain meteorological environmental parameters and data acquisition rate for the current flight phase. Based on the correlation between the meteorological environmental parameters and the statistical characteristics of quality scores in historical flight phases, determine the quantitative relationship of the influence of meteorological environmental conditions on the quality distribution of cloud particle images, and calculate the expected quality offset under the current meteorological environmental conditions.

[0086] Based on the data acquisition rate, the available capacity of onboard storage resources, and the minimum requirement for the number of valid images, the target data retention rate under the current acquisition state is calculated by combining storage capacity constraints and data requirement constraints.

[0087] Based on the quantile distribution curve in the quality score statistical features, the quantile distribution curve is corrected by combining the expected quality offset;

[0088] Find the quantiles on the corrected quantile distribution curve that correspond to the target data retention rate, and determine the quality score of the quantiles as the quality screening threshold.

[0089] During aircraft-based cloud particle detection, a large number of cloud particle images need to be collected for analysis and research. However, due to the complex and variable flight environment, the quality of the raw cloud particle images varies greatly. Saving all images without filtering would not only consume a large amount of storage space but also increase the difficulty of post-processing. To address this issue, firstly, the quality scores of all images in the raw cloud particle image sequence are statistically analyzed to obtain the statistical characteristics of the quality scores. The quality score can be calculated based on multiple dimensions such as image sharpness, contrast, brightness, and particle integrity, and its value ranges from 0 to 1, where 1 represents the highest quality. By statistically analyzing the quality scores of all images, statistical characteristics such as the mean, standard deviation, median, and quantile distribution of the scores are obtained. For example, in one flight mission, 10,000 cloud particle images were collected, with a mean quality score of 0.72, a standard deviation of 0.15, and a median of 0.75, and the scores exhibit an approximately normal distribution. These statistical characteristics provide the basic data for subsequently determining the filtering threshold.

[0090] Next, the meteorological environmental parameters and data acquisition rate for the current flight phase are obtained. These parameters include atmospheric temperature, humidity, air pressure, wind speed, and particle concentration, which are collected in real time by onboard sensors. The data acquisition rate refers to the number of cloud particle images acquired per unit time, and is affected by the performance of the detection equipment and the particle concentration. For example, when the aircraft is flying in a high-altitude, low-temperature region, with a temperature of -40°C, relative humidity of 85%, and a cloud particle concentration of 200 particles / cm³, the situation may be different. 3 At that time, the data acquisition rate reached 50 images per second.

[0091] Based on the correlation between meteorological environmental parameters and the statistical characteristics of quality scores in historical flight phases, the quantitative relationship of the impact of meteorological environmental conditions on the quality distribution of cloud particle images is determined. A mathematical model of meteorological parameters and quality distribution shift is established by analyzing historical data. For example, for every 10°C decrease in temperature, the average quality score decreases by 0.05; for every 10% increase in relative humidity, the standard deviation of the quality score increases by 0.02. Based on these relationships, the expected quality shift under current meteorological environmental conditions is calculated. If the current temperature is 20°C lower than the historical average temperature and the humidity is 15% higher than the historical average humidity, then the expected quality shift is -0.1 + 0.03 = -0.07.

[0092] Based on the data acquisition rate, available onboard storage capacity, and minimum requirements for the number of valid images, the target data retention rate under the current acquisition conditions is calculated. Assuming the onboard storage system has 200GB of available space, each image occupies an average of 2MB, the current acquisition rate is 50 images / second, and the estimated remaining flight time is 2 hours, a total of approximately 360,000 images will be acquired, occupying approximately 720GB. However, the storage capacity only allows for storing approximately 100,000 images; therefore, the target data retention rate should be set at approximately 28%. Furthermore, if research requirements necessitate retaining at least 20,000 valid images, the target retention rate should not be lower than 5.6%. Considering both constraints, the target data retention rate is determined to be 28%.

[0093] Based on the quantile distribution curve in the statistical features of the quality score obtained in the preceding steps, the quantile distribution curve is corrected by incorporating the expected quality offset. If the original quantile distribution shows that the quality score corresponds to the 75th percentile at 0.85, considering the -0.07 offset caused by the current environment, the distribution curve is shifted to the left by 0.07, and after correction, the 75th percentile corresponds to 0.78.

[0094] Finally, the quantile corresponding to the target data retention rate is found on the corrected quantile distribution curve. The quality score of this quantile is determined as the quality screening threshold. For a target retention rate of 28%, the corresponding quantile is 72%. The corrected quantile distribution curve shows that the quality score corresponding to this quantile is 0.76. Therefore, 0.76 is set as the quality screening threshold for the current acquisition state, and only cloud particle images with a quality score higher than this threshold are retained.

[0095] The above methods enable dynamic threshold adjustment based on the current flight environment and resource constraints, ensuring data quality while meeting storage constraints and improving the efficiency and quality of cloud particle data acquisition. This method is applicable to various aviation meteorological observation missions, intelligently adjusting data filtering strategies according to real-time conditions, and providing high-quality basic data for subsequent scientific research.

[0096] In one optional implementation, meteorological environmental parameters and data acquisition rate for the current flight phase are acquired. Based on the correlation between the meteorological environmental parameters and the statistical characteristics of quality scores in historical flight phases, the quantitative relationship of the influence of meteorological environmental conditions on cloud particle image quality distribution is determined. The expected quality offset under the current meteorological environmental conditions is calculated, including:

[0097] Meteorological environmental parameters and data acquisition rates for the current flight phase are obtained from airborne meteorological sensors and flight control systems, and meteorological environmental parameter records and corresponding quality score statistical feature records for multiple historical flight phases are extracted from the historical flight database.

[0098] For each historical flight phase, a functional mapping relationship is determined between the meteorological environmental parameter records and the quality score statistical feature records. A quantitative expression for the influence of meteorological environmental conditions on the quality score distribution is then determined by fitting the functional mapping relationship.

[0099] Based on the meteorological environmental parameters of the current flight phase and the quantitative relationship expression, the expected quality score distribution characteristics under the current meteorological environmental conditions are determined.

[0100] The deviation between the expected quality score distribution characteristics and the actual quality score statistical characteristics of the current flight phase is determined as the expected quality offset.

[0101] The expected mass offset is determined based on the ratio between the data acquisition rate of the current flight phase and the data acquisition rate of the historical flight phase.

[0102] Figure 2 This is a schematic diagram illustrating the process of determining the expected quality offset according to an embodiment of the present invention. Figure 2 As shown, the meteorological environmental parameters and data acquisition rate for the current flight phase can be obtained through airborne meteorological sensors and the flight control system. The airborne meteorological sensors include temperature sensors, humidity sensors, barometric pressure sensors, and wind speed and direction sensors, used to collect meteorological parameters such as temperature, humidity, barometric pressure, and wind speed and direction in the flight environment in real time. The flight control system provides flight status parameters such as flight altitude, flight speed, and flight attitude, while simultaneously recording the data acquisition rate of the cloud particle imaging device, i.e., the number of cloud particle images acquired per second. These parameters are transmitted to the central processing unit via a data bus, forming a set of meteorological environmental parameters and a data acquisition rate record for the current flight phase.

[0103] Meteorological environmental parameter records and corresponding quality score statistical feature records for multiple historical flight phases were extracted from the historical flight database. The database stores meteorological environmental parameters and corresponding cloud particle image quality score statistical features from different flight phases in past missions, including statistical measures such as the mean, variance, skewness, and kurtosis of the quality scores. The extraction operation is based on key parameters such as current flight altitude, temperature, and humidity, selecting historical flight records under similar meteorological conditions to ensure a high correlation between the extracted historical data and the current flight environment.

[0104] For each historical flight phase, the functional mapping relationship between meteorological environmental parameter records and quality score statistical characteristic records was determined. First, the meteorological environmental parameters were standardized to eliminate the influence of dimensional differences between different parameters. Then, multivariate regression analysis was used, with meteorological environmental parameters as independent variables and quality score statistical characteristics as dependent variables, to establish the functional mapping relationship. Optimization algorithms such as least squares or gradient descent were used to solve the coefficients of the regression equation, yielding the mapping function between meteorological environmental parameters and quality score statistical characteristics.

[0105] The quantitative relationship expression for the influence of meteorological environmental conditions on the distribution of quality scores is determined by fitting a function mapping relationship. During the fitting process, methods such as multinomial regression, support vector regression, or neural networks are employed to establish a nonlinear mapping model between meteorological environment and quality score distribution characteristics for different combinations of meteorological parameters. After model training, model performance is evaluated through cross-validation, and the model with the strongest generalization ability is selected as the final quantitative relationship expression. This expression can accurately predict the distribution characteristics of cloud particle image quality scores based on input meteorological environmental parameters.

[0106] Based on the meteorological environmental parameters and quantitative relationship expressions of the current flight phase, the expected distribution characteristics of the quality score under the current meteorological environmental conditions are determined. The meteorological environmental parameters collected during the current flight phase are substituted into the established quantitative relationship expressions to calculate the expected distribution characteristics of the cloud particle image quality score under the current meteorological conditions, including statistical quantities such as the expected mean and expected variance. These expected distribution characteristics reflect the potential impact of the current meteorological environmental conditions on the quality of cloud particle images.

[0107] The deviation between the expected quality score distribution characteristics and the actual quality score statistical characteristics during the current flight phase is defined as the expected quality offset. The actual quality score statistical characteristics during the current flight phase are obtained by quality assessment of real-time acquired cloud particle images, with assessment indicators including image sharpness, contrast, and signal-to-noise ratio. The Euclidean distance or Mahalanobis distance between the actual statistical characteristics and the expected distribution characteristics is calculated as the initial estimate of the expected quality offset.

[0108] Based on the ratio between the data acquisition rate of the current flight phase and the data acquisition rate of the historical flight phases, the expected quality offset is further optimized. Differences in data acquisition rates affect the consistency and stability of image quality, therefore, the initial offset needs to be corrected. Specifically, if the current acquisition rate is higher than the historical acquisition rate, the image quality is negatively affected by the compression of the acquisition time; conversely, if the current acquisition rate is lower than the historical acquisition rate, a higher quality image is obtained. By calculating the ratio of the current to the historical acquisition rate and combining it with empirical correction coefficients, the expected quality offset is weighted and adjusted to obtain the final expected quality offset.

[0109] In a practical application scenario, during an atmospheric science observation flight mission, the aircraft flew at an altitude of 10,000 meters, a temperature of -40 degrees Celsius, and a relative humidity of 35%. The cloud particle imaging device acquired data at a rate of 50 frames per second. By querying the historical flight database, five historical flight records under similar meteorological conditions were selected. The average data acquisition rate of these records was 40 frames per second. Based on the quantitative relationship expression established from the historical data, the predicted average cloud particle image quality score under the current meteorological conditions should be 85 points (out of 100). However, the actual measured average quality score was 78 points. Considering that the current acquisition rate is 25% higher than the historical average, a correction calculation yielded a quality expectation offset of -5 points, indicating that the current meteorological conditions negatively impacted image quality, necessitating adjustments to the imaging parameters to improve image quality.

[0110] The expected mass offset obtained through the above steps provides an important basis for subsequent cloud particle image quality assessment and imaging parameter adjustment, effectively improving the accuracy and reliability of atmospheric science observation data.

[0111] In one optional implementation, images in the original cloud particle image sequence with quality scores higher than the quality screening threshold are marked as valid images and retained, while images with quality scores lower than the quality screening threshold are marked as invalid images and removed, resulting in a filtered image set, including:

[0112] The quality score of each image in the original cloud particle image sequence is compared with the quality screening threshold, and a screening label for each image is determined based on the comparison result.

[0113] Identify boundary images where the absolute value of the difference between the quality score and the quality screening threshold is less than a preset boundary range, and extract key quality defect type information from the boundary images.

[0114] Determine whether the critical quality defect type belongs to the set of tolerable defect types, and adjust the filtering markers of the boundary image based on the determination result;

[0115] The time interval between adjacent images that are screened as valid images is detected. When the time interval exceeds a continuity threshold, the candidate compensation image with the highest quality score is searched among the invalid images located between the adjacent images in the original cloud particle image sequence.

[0116] Determine whether the quality score of the candidate compensated image meets the downgrade retention condition. If it does, correct the screening mark of the candidate compensated image to a valid image mark. Organize all images with the screening mark as valid image marks in the order of acquisition time to obtain the filtered image set.

[0117] When performing quality screening on the original cloud particle image sequence, the first step is to calculate a quality score for each image and compare it with a preset quality screening threshold to determine whether the image should be retained. This process considers not only the quality characteristics of the image itself but also the temporal continuity of the image sequence.

[0118] The quality score of each image in the original cloud particle image sequence is compared with a quality screening threshold. The comparison result is used to determine the screening label for each image. Specifically, when the quality score of an image is higher than the quality screening threshold, the image is marked as a valid image; when the quality score is lower than the quality screening threshold, the image is marked as an invalid image. The quality score can be calculated by comprehensively considering multiple indicators such as image sharpness, contrast, and signal-to-noise ratio. For example, a weighted average method can be used, where each indicator is assigned different weights according to its importance and then summed.

[0119] During the comparison process, special attention is paid to images where the absolute value of the difference between the quality score and the quality screening threshold is less than a preset boundary range; these images are defined as boundary images. The preset boundary range can be set to 5% to 10% of the quality screening threshold. For boundary images, it is necessary to further extract their key quality defect type information, which includes, but is not limited to, specific defect types such as image blurriness, abnormal brightness, and insufficient contrast.

[0120] After extracting the key quality defect type information, it is determined whether the defect type belongs to the tolerable defect type set. The tolerable defect type set refers to defect types that have a minor impact on subsequent processing, such as slight brightness unevenness or small noise in edge areas. If the key quality defect type of the boundary image belongs to the tolerable defect type set, the filtering label of the boundary image is corrected to a valid image label; otherwise, it remains an invalid image label.

[0121] After initial screening, the time interval between adjacent images marked as valid is detected. This time interval can be calculated from the timestamps in the image metadata. When the time interval between adjacent valid images exceeds a preset continuity threshold, compensation processing is required. The continuity threshold can be set according to the specific needs of cloud particle observation, for example, twice the normal acquisition frequency.

[0122] If the time interval exceeds a continuity threshold, the image with the highest quality score is searched among all invalid images located between two adjacent valid images in the original cloud particle image sequence, and this image is selected as a candidate compensation image. By comparing the quality scores of all invalid images and selecting the image with the highest score for compensation, the quality of the compensation image can be guaranteed to the maximum extent.

[0123] The quality score of the candidate compensated image is evaluated to determine whether it meets the downgrade retention criteria. The downgrade retention criteria can be set to be lower than the original quality screening threshold but higher than the downgrade threshold. The downgrade threshold is usually set to 80% to 90% of the quality screening threshold. If the candidate compensated image meets the downgrade retention criteria, its screening label is corrected to a valid image label, thereby making up for the missing data in the time series.

[0124] Finally, all images marked as valid images are organized in chronological order of acquisition time to form a filtered image set. This processed image set ensures that the image quality meets the basic requirements while maintaining the continuity of the time series, which is beneficial for subsequent analysis and processing of cloud particle image sequences.

[0125] In practical applications, parameters such as the quality screening threshold, preset boundary range, set of tolerable defect types, and continuity threshold can be adjusted according to specific observation tasks. For example, for tasks that require precise analysis of cloud particle morphology characteristics, the quality screening threshold can be increased; while for tasks that mainly focus on the changing trends in the number of cloud particles, the quality requirements can be appropriately relaxed, with greater emphasis placed on the continuity of the time series.

[0126] This multi-level, multi-dimensional screening strategy can preserve the effective information in the time series to the maximum extent while ensuring the quality of cloud particle images, thus providing a reliable data foundation for subsequent cloud particle characteristic analysis.

[0127] In one optional implementation, the quality feature distribution pattern of valid images and the defect feature clustering result of invalid images in the filtered image set are determined. Based on the quality feature distribution pattern and the defect feature clustering result, the feature weight coefficients and classification decision boundary in the multi-task learning architecture are updated, including:

[0128] Based on the comprehensive quality feature vector of all valid images in the filtered image set, calculate the quality feature distribution pattern of the valid images;

[0129] The defect feature clustering results of the invalid images are determined based on the key quality defect types and comprehensive quality feature vectors corresponding to all invalid images in the original cloud particle image sequence;

[0130] For each feature dimension, the minimum distance between the mean in the quality feature distribution pattern and the feature center of each defect category in the defect feature clustering result is calculated, and the feature discrimination index is calculated based on the minimum distance and the quality feature distribution pattern.

[0131] The feature weight coefficients of the corresponding feature dimensions in the multi-task learning architecture are adjusted according to the value of the feature discrimination index, so that the feature weight coefficients are positively correlated with the feature discrimination index.

[0132] Based on the feature center location and feature dispersion of each defect category in the defect feature clustering results, the classification decision boundary corresponding to the defect category is adjusted in the defect recognition decision layer of the multi-task learning architecture so that the classification decision boundary matches the actual feature distribution range of the defect category.

[0133] In image quality screening applications, in order to dynamically update the feature weights and decision boundaries in the multi-task learning architecture, it is necessary to optimize based on the quality feature distribution pattern of effective images and the defect feature clustering results of invalid images. The following details the specific implementation process of updating feature weight coefficients and classification decision boundaries.

[0134] First, based on the comprehensive quality feature vector of all valid images in the filtered image set, the quality feature distribution pattern of the valid images is calculated. Specifically, assuming the filtered valid image set contains N images, and the comprehensive quality feature vector of each image has dimension M, the mean μj and standard deviation σj are calculated for each feature dimension j. The mean μj is obtained by averaging the feature values ​​of all N images in that dimension, and the standard deviation σj is obtained by calculating the dispersion of the feature values ​​in that dimension. This yields the distribution characteristics of the valid images across each feature dimension, forming a quality feature distribution pattern P, which includes the mean and standard deviation information for each feature dimension.

[0135] Next, based on the key quality defect types and comprehensive quality feature vectors corresponding to all invalid images in the original cloud particle image sequence, the defect feature clustering results of the invalid images are determined. Assuming there are K defect types, for each defect type k, all invalid images labeled as that defect type are collected, and their feature centers ck (i.e., the mean of each feature dimension) and the dispersion dk of that defect type on each feature dimension are calculated (which can be obtained by calculating the standard deviation or interquartile range). Thus, for each defect type k, there is a feature center ck and a dispersion dk, constituting the defect feature clustering result C.

[0136] Then, for each feature dimension j, the minimum distance value distj between the mean μj in the quality feature distribution pattern and the feature centers of each defect category in the defect feature clustering result is calculated. Specifically, for feature dimension j, the distance between μj and the center ck[j] of each defect type is calculated, and the minimum value is taken as distj. The calculation formula is: distj = min(|μj - ck[j]|), where k takes values ​​from 1 to K.

[0137] Based on the minimum distance value distj and the quality feature distribution pattern, the feature discrimination index Dj is calculated. The feature discrimination index measures the ability of feature dimension j to distinguish between valid and invalid images. The calculation formula is: Dj = distj / (σj + ε), where ε is a small constant used to avoid a zero denominator. When distj is large and σj is small, it indicates that the feature dimension has high discrimination ability because the distribution of valid images along this dimension is concentrated and far from the distribution center of defect types.

[0138] Based on the value of the feature discrimination index Dj, the feature weight coefficients wj for the corresponding feature dimensions in the multi-task learning architecture are adjusted. The weight coefficients wj and the feature discrimination index Dj are positively correlated. They can be directly assigned after normalization, or transformed using a mapping function f, for example: wj = f(Dj), where f can be a sigmoid function, a softmax function, or a simple linear mapping. In practical applications, an upper limit wmax and a lower limit wmin can be set for the weights to ensure that the adjusted weights are within a reasonable range, avoiding excessively large or small weights for certain feature dimensions.

[0139] Finally, based on the feature center position ck and feature dispersion dk of each defect category k in the defect feature clustering results, the classification decision boundary corresponding to the defect category is adjusted in the defect recognition decision layer of the multi-task learning architecture. For the case of using a linear classifier, the normal vector and bias term of the decision hyperplane can be adjusted; for the distance-based classifier, the representative point position and threshold distance of each category can be adjusted.

[0140] Taking Support Vector Machine (SVM) as an example, new support vectors can be constructed based on the defect feature centers ck and their dispersion dk, and the decision hyperplane can be adjusted. If the feature dispersion dk of defect type k is large, it indicates that the feature distribution of this type is relatively dispersed, and the fault tolerance space of the decision boundary should be expanded; conversely, if dk is small, it indicates that the feature distribution of this type is concentrated, and the decision boundary can be tightened to improve the recognition accuracy.

[0141] Through the above steps, the feature weight coefficients and classification decision boundaries in the multi-task learning architecture are dynamically updated, enabling it to better adapt to the feature distribution characteristics of the current image set and improve the accuracy and efficiency of cloud particle image quality screening. The updated multi-task learning architecture can more accurately distinguish between valid images and various types of invalid images, achieving refined quality control of cloud particle images.

[0142] In practical applications, the above update process can be performed periodically, or an update can be triggered when a significant change in image distribution characteristics is detected, in order to maintain continuous adaptability and efficiency.

[0143] In one optional implementation, the defect feature clustering result of the invalid images is determined based on the key quality defect types and comprehensive quality feature vectors corresponding to all invalid images in the original cloud particle image sequence, including:

[0144] Traverse all invalid images in the original cloud particle image sequence, extract the key quality defect types corresponding to each invalid image, count the frequency of occurrence of each key quality defect type, and determine the set of all defect types existing in the original cloud particle image sequence;

[0145] For each defect type in the set of all defect types, extract the comprehensive quality feature vectors corresponding to all invalid images of the original cloud particle image sequence whose key quality defect type is that defect type, and obtain the feature vector set corresponding to the defect type;

[0146] For each defect type, a feature center is calculated based on the comprehensive quality feature vectors in the feature vector set.

[0147] For each defect type, calculate the distance between each comprehensive quality feature vector in the feature vector set corresponding to the defect type and the feature center of the defect type, and determine the feature dispersion of the defect type based on the distance value;

[0148] The feature center and feature dispersion corresponding to each defect type in the set of all defect types are used as the defect feature clustering result.

[0149] When determining the clustering results of defect features of invalid images based on the key quality defect types and comprehensive quality feature vectors corresponding to all invalid images in the original cloud particle image sequence, it is first necessary to traverse all invalid images in the original cloud particle image sequence. For each invalid image, its corresponding key quality defect type is extracted, such as insufficient sharpness, abnormal contrast, abnormal brightness, and blurred edges. During the extraction process, each image can be analyzed using a preset image quality assessment model, which identifies quality problems in the image based on image processing technology. The frequency of occurrence of each key quality defect type is statistically analyzed. For example, in a sequence containing one thousand invalid cloud particle images, three hundred images have insufficient sharpness, two hundred images have abnormal contrast, four hundred images have abnormal brightness, and one hundred images have blurred edges. Through this statistical analysis, the set of all defect types present in the original cloud particle image sequence is determined.

[0150] For each defect type in the entire defect type set, it is necessary to extract the comprehensive quality feature vector corresponding to all invalid images in the original cloud particle image sequence whose key quality defect type is that defect type. The comprehensive quality feature vector is a multi-dimensional feature set characterizing image quality, which can include multiple aspects such as image sharpness, brightness, contrast, texture features, and edge features. For example, for the defect type of insufficient sharpness, the comprehensive quality feature vectors of all invalid images marked as having insufficient sharpness are extracted to form a feature vector set. These feature vectors can be calculated using image processing algorithms, such as calculating image sharpness using the Laplacian operator, calculating brightness and contrast using histogram analysis, and calculating texture features using the gray-level co-occurrence matrix, etc.

[0151] For each defect type, a feature vector set is used to calculate the feature center of the defect type based on the comprehensive quality feature vectors in that set. The feature center can be calculated using the mean center method, which involves averaging the values ​​of each dimension of all feature vectors under the same defect type to obtain a central vector that represents the typical characteristics of that type of defect. For example, for the feature vector set of the insufficient sharpness type, the average value of all feature vectors in each dimension is calculated to form the feature center of the insufficient sharpness type.

[0152] For each defect type, the distance between each comprehensive quality feature vector in the feature vector set corresponding to that defect type and the feature center of that defect type is calculated. Distance calculation can employ methods such as Euclidean distance, Manhattan distance, and cosine similarity, selecting the appropriate distance metric for the specific application scenario. Based on the calculated distance values, the feature dispersion of the defect type is determined. Feature dispersion represents the degree of dispersion of samples within the same defect type. Dispersion can be measured by calculating the standard deviation or variance of all distance values. Smaller dispersion indicates more consistent feature performance for that type of defect, while larger dispersion indicates greater variation in feature performance.

[0153] Finally, the feature centers and feature dispersions corresponding to each defect type in the entire defect type set are used as the defect feature clustering results. These clustering results can be used for subsequent image quality analysis and processing. For example, for a new cloud particle image, the distance between its comprehensive quality feature vector and the feature centers of each defect type can be calculated, and combined with the feature dispersion, to determine whether the image has a specific type of defect and the severity of the defect.

[0154] In practical applications, defect feature clustering results can be used to guide parameter adjustments for image acquisition equipment. For example, if a large number of images are found to have abnormal brightness and the feature dispersion is small, it indicates that this type of defect has stable feature characteristics and is caused by improper equipment parameter settings. In this case, the brightness parameters of the acquisition equipment can be adjusted. If the feature dispersion of a certain type of defect is large, it indicates that this type of defect has diverse manifestations and requires more complex processing strategies or further subdivision of defect types.

[0155] The above methods enable a systematic analysis of the quality defect characteristics of invalid images in the original cloud particle image sequence, providing strong support for subsequent image quality control and improvement.

[0156] The airborne cloud particle image quality intelligent assessment and screening system of this invention includes:

[0157] The first unit is used to perform multi-scale spatial domain and frequency domain joint analysis on each image in the original cloud particle image sequence acquired by the airborne cloud particle detection equipment, and extract the comprehensive quality feature vector;

[0158] The second unit is used to simultaneously calculate the quality score and key quality defect types of each image based on the comprehensive quality feature vector using a multi-task learning architecture;

[0159] The third unit is used to calculate a quality screening threshold that matches the current data acquisition status based on the statistical distribution characteristics of the quality score, combined with the meteorological environmental parameters and data acquisition rate of the current flight phase.

[0160] The fourth unit is used to mark images in the original cloud particle image sequence with a quality score higher than the quality screening threshold as valid images and retain them, and to mark images with a quality score lower than the quality screening threshold as invalid images and remove them, thereby obtaining a set of filtered images;

[0161] The fifth unit is used to determine the quality feature distribution pattern of valid images and the defect feature clustering result of invalid images in the filtered image set, and to update the feature weight coefficients and classification decision boundary in the multi-task learning architecture based on the quality feature distribution pattern and the defect feature clustering result.

[0162] A third aspect of the present invention provides an electronic device, comprising:

[0163] processor;

[0164] Memory used to store processor-executable instructions;

[0165] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0166] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0167] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0168] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent assessment and screening method for airborne cloud particle image quality, characterized in that, include: Multi-scale spatial and frequency domain joint analysis was performed on each image in the raw cloud particle image sequence acquired by the airborne cloud particle detection equipment to extract the comprehensive quality feature vector; Based on the comprehensive quality feature vector, a multi-task learning architecture is used to simultaneously calculate the quality score and key quality defect types for each image; Based on the statistical distribution characteristics of the quality score, and combined with the meteorological environmental parameters and data acquisition rate of the current flight phase, a quality screening threshold matching the current acquisition status is calculated; Images in the original cloud particle image sequence with a quality score higher than the quality screening threshold are marked as valid images and retained, while images with a quality score lower than the quality screening threshold are marked as invalid images and removed, resulting in a filtered image set. Determine the quality feature distribution pattern of valid images and the defect feature clustering result of invalid images in the filtered image set. Based on the quality feature distribution pattern and the defect feature clustering result, update the feature weight coefficients and classification decision boundary in the multi-task learning architecture. The step of calculating a quality screening threshold that matches the current data acquisition status based on the statistical distribution characteristics of the quality score, combined with the meteorological environmental parameters and data acquisition rate of the current flight phase, includes: Statistical analysis was performed on the quality scores of all images in the original cloud particle image sequence to obtain the statistical characteristics of the quality scores; Obtain meteorological environmental parameters and data acquisition rate for the current flight phase. Based on the correlation between the meteorological environmental parameters and the statistical characteristics of quality scores in historical flight phases, determine the quantitative relationship of the influence of meteorological environmental conditions on the quality distribution of cloud particle images, and calculate the expected quality offset under the current meteorological environmental conditions. Based on the data acquisition rate, the available capacity of onboard storage resources, and the minimum requirement for the number of valid images, the target data retention rate under the current acquisition state is calculated by combining storage capacity constraints and data requirement constraints. Based on the quantile distribution curve in the quality score statistical features, the quantile distribution curve is corrected by combining the expected quality offset; Find the quantiles on the corrected quantile distribution curve that correspond to the target data retention rate, and determine the quality score of the quantiles as the quality screening threshold.

2. The method of claim 1, wherein, Multi-scale spatial and frequency domain joint analysis was performed on each image in the raw cloud particle image sequence acquired by the airborne cloud particle detection equipment to extract a comprehensive quality feature vector, including: Each image in the original cloud particle image sequence is subjected to multi-scale spatial decomposition to obtain spatial domain quality feature sub-vectors; Perform a frequency domain transformation on each image in the original cloud particle image sequence to obtain a frequency domain quality feature subvector; Based on the high-frequency gradient components in the spatial domain quality feature vector and the high-frequency energy distribution characteristics in the frequency domain quality feature vector, particle edge information and noise interference components in the image are distinguished, and a noise intensity assessment value characterizing the noise level is calculated. Based on the edge continuity metric in the spatial domain quality feature sub-vector, and combined with the noise intensity evaluation value, the edge detection results are weighted for reliability. By constructing the connectivity topology of edge pixels and the consistency constraint of edge intensity, the edge quality evaluation value characterizing the integrity of particle edges is calculated. The spatial domain quality feature vector, the frequency domain quality feature vector, the noise intensity assessment value, and the edge quality assessment value are weighted and combined to obtain the comprehensive quality feature vector.

3. The method of claim 1, wherein, Obtain meteorological environmental parameters and data acquisition rate for the current flight phase. Based on the correlation between the meteorological environmental parameters and the statistical characteristics of quality scores in historical flight phases, determine the quantitative relationship of the impact of meteorological environmental conditions on cloud particle image quality distribution. Calculate the expected quality offset under the current meteorological environmental conditions, including: Meteorological environmental parameters and data acquisition rates for the current flight phase are obtained from airborne meteorological sensors and flight control systems, and meteorological environmental parameter records and corresponding quality score statistical feature records for multiple historical flight phases are extracted from the historical flight database. For each historical flight phase, a functional mapping relationship is determined between the meteorological environmental parameter records and the quality score statistical feature records. A quantitative expression for the influence of meteorological environmental conditions on the quality score distribution is then determined by fitting the functional mapping relationship. Based on the meteorological environmental parameters of the current flight phase and the quantitative relationship expression, the expected quality score distribution characteristics under the current meteorological environmental conditions are determined. The deviation between the expected quality score distribution characteristics and the actual quality score statistical characteristics during the current flight phase is defined as the expected quality offset. The expected mass offset is determined based on the ratio between the data acquisition rate of the current flight phase and the data acquisition rate of the historical flight phase.

4. The method of claim 1, wherein, Images in the original cloud particle image sequence with quality scores higher than the quality screening threshold are marked as valid images and retained, while images with quality scores lower than the quality screening threshold are marked as invalid images and removed, resulting in a filtered image set, including: The quality score of each image in the original cloud particle image sequence is compared with the quality screening threshold, and a screening label for each image is determined based on the comparison result. Identify boundary images where the absolute value of the difference between the quality score and the quality screening threshold is less than a preset boundary range, and extract key quality defect type information from the boundary images. Determine whether the critical quality defect type belongs to the tolerable defect type set, and adjust the filtering markers of the boundary image based on the determination result; The time interval between adjacent images that are screened as valid images is detected. When the time interval exceeds a continuity threshold, the candidate compensation image with the highest quality score is searched among the invalid images located between the adjacent images in the original cloud particle image sequence. Determine whether the quality score of the candidate compensated image meets the downgrade retention condition. If it does, correct the screening mark of the candidate compensated image to a valid image mark. Organize all images with the screening mark as valid image marks in the order of acquisition time to obtain the filtered image set.

5. The method of claim 1, wherein, Determine the quality feature distribution pattern of valid images and the defect feature clustering result of invalid images in the filtered image set. Based on the quality feature distribution pattern and the defect feature clustering result, update the feature weight coefficients and classification decision boundary in the multi-task learning architecture, including: Based on the comprehensive quality feature vector of all valid images in the filtered image set, calculate the quality feature distribution pattern of the valid images; The defect feature clustering results of the invalid images are determined based on the key quality defect types and comprehensive quality feature vectors corresponding to all invalid images in the original cloud particle image sequence; For each feature dimension, the minimum distance between the mean in the quality feature distribution pattern and the feature center of each defect category in the defect feature clustering result is calculated, and the feature discrimination index is calculated based on the minimum distance and the quality feature distribution pattern. The feature weight coefficients of the corresponding feature dimensions in the multi-task learning architecture are adjusted according to the value of the feature discrimination index, so that the feature weight coefficients are positively correlated with the feature discrimination index. Based on the feature center location and feature dispersion of each defect category in the defect feature clustering results, the classification decision boundary corresponding to the defect category is adjusted in the defect recognition decision layer of the multi-task learning architecture so that the classification decision boundary matches the actual feature distribution range of the defect category.

6. The method of claim 5, wherein, The defect feature clustering results of the invalid images are determined based on the key quality defect types and comprehensive quality feature vectors corresponding to all invalid images in the original cloud particle image sequence, including: Traverse all invalid images in the original cloud particle image sequence, extract the key quality defect types corresponding to each invalid image, count the frequency of occurrence of each key quality defect type, and determine the set of all defect types existing in the original cloud particle image sequence; For each defect type in the set of all defect types, extract the comprehensive quality feature vectors corresponding to all invalid images of the original cloud particle image sequence whose key quality defect type is that defect type, and obtain the feature vector set corresponding to the defect type; For each defect type, a feature center is calculated based on the comprehensive quality feature vectors in the feature vector set. For each defect type, calculate the distance between each comprehensive quality feature vector in the feature vector set corresponding to the defect type and the feature center of the defect type, and determine the feature dispersion of the defect type based on the distance value; The feature center and feature dispersion corresponding to each defect type in the set of all defect types are used as the defect feature clustering result.

7. An airborne cloud particle image quality intelligent assessment and screening system, used to implement the method as described in any one of claims 1-6, characterized in that, include: The first unit is used to perform multi-scale spatial domain and frequency domain joint analysis on each image in the original cloud particle image sequence acquired by the airborne cloud particle detection equipment, and extract the comprehensive quality feature vector; The second unit is used to simultaneously calculate the quality score and key quality defect types of each image based on the comprehensive quality feature vector using a multi-task learning architecture; The third unit is used to calculate a quality screening threshold that matches the current data acquisition status based on the statistical distribution characteristics of the quality score, combined with the meteorological environmental parameters and data acquisition rate of the current flight phase. The fourth unit is used to mark and retain images in the original cloud particle image sequence whose quality score is higher than the quality screening threshold, and to mark and remove images whose quality score is lower than the quality screening threshold, thereby obtaining a set of filtered images. The fifth unit is used to determine the quality feature distribution pattern of valid images and the defect feature clustering result of invalid images in the filtered image set, and to update the feature weight coefficients and classification decision boundary in the multi-task learning architecture based on the quality feature distribution pattern and the defect feature clustering result.

8. An electronic device, comprising: include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having stored thereon computer program instructions, wherein, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images