Image quality evaluation system based on multiple filters and deep learning

Abstract

The invention relates to an image quality evaluation system based on multiple filters and deep learning, which comprises an image input system, a multi-filter evaluation module, a quality grading module, a deep learning model, an image storage system and an inference pipeline, and is characterized in that a prediction model is trained by utilizing an accumulated data set of images with corresponding filter values and quality labels; these models may learn relationships between visual image features and quality measurements. The system is enabled to develop automatic quality assessment capabilities, replication of assessment results of a multi-filter assessment module, without explicit filter calculations during subsequent reasoning operations, generation of predictive algorithms, integration of image processing and quality assessment through interconnection module operations arranged in a structured architecture, and improvement of quality assessment efficiency. Consecutive quality scores are generated rather than binary classification, achieving a detailed quality assessment in the lineage of image features and application requirements.

Description

TECHNICAL FIELD

[0001] The application relates to an image quality evaluation system, in particular to an image quality evaluation system based on multiple filters and deep learning applied to the field of image processing. BACKGROUND

[0002] Image quality evaluation has become an essential component of modern digital imaging systems, applied in many fields, from consumer photography and video streaming to medical imaging and industrial inspection, when digital images are processed, transmitted and stored through various pipelines, they often degrade due to compression artifacts, noise, blur and other distortions, which affect their perceptual quality and utility, traditional image quality evaluation methods usually rely on reference-based methods, comparing the processed image with the original, undistorted version, however, in many real-world scenarios, reference images are not available, which makes no-reference image quality evaluation methods more practical and ideal, these methods must evaluate image quality based only on the features of the image itself, without accessing the original reference image.

[0003] Chinese patent CN113658130B discloses a no-reference screen content image quality evaluation method based on a double twin network, which is beneficial to use the shallow significant LBP structural features that can be extracted by traditional methods, and improves the performance of no-reference screen content image quality evaluation, Chinese patent CN111047618B discloses a no-reference screen content image quality evaluation method based on multi-scale, by comprehensively considering the significant features of screen content images and human visual system, extracting features with stronger representation ability for distortion information of distorted images, significantly improving the performance of no-reference screen content image quality evaluation method, suitable for screen content image quality evaluation of various distortion types and different distortion degrees, the predicted quality score of the distorted screen content image is closer to the subjective score.

[0004] Existing no-reference image quality evaluation systems face several limitations, many current methods rely on single-index evaluation, such as measuring sharpness or clarity only through techniques such as Laplacian variance calculation, although these methods can provide basic quality indicators, they often prove to be insufficient when performing robust evaluation under various conditions including different lighting environments, noise levels and motion blur scenarios. SUMMARY

[0005] In view of the above prior art, the technical problem to be solved by the present application is that the existing no-reference image quality evaluation system faces several limitations, many current methods rely on single-index evaluation, which is often insufficient.

[0006] To solve the above problems, the application provides an image quality evaluation system based on multi-filter and deep learning, which comprises an image input system, a multi-filter evaluation module, a quality grading module, a deep learning model, an image storage system and an inference pipeline.

[0007] The multi-filter evaluation module comprises a Laplacian variance filter, an FFT variance filter and a gradient mean filter.

[0008] The quality grading module comprises a voting average grade (VAG) calculator and a weighted average grade (WAG) calculator.

[0009] The deep learning model comprises a visual transformer encoder and a fully connected network connected to each other.

[0010] The inference pipeline comprises a quality score predictor and an image selector.

[0011] In the above image quality evaluation system based on multi-filter and deep learning, the interconnection module operation arranged in the structured architecture is used for comprehensive image processing and quality evaluation, and a continuous quality score instead of a binary classification is generated to realize detailed quality evaluation in the spectrum of image features and application requirements.

[0012] As a further improvement of the present application, the image quality evaluation and data generation method mainly comprises:

[0013] S1, input the image to be evaluated through the image input system, and after gray processing of the image, apply it to the multi-filter evaluation module, and calculate the Laplacian variance by detecting the rapid intensity change in the gray image, calculate the FFT variance by applying 2D fast Fourier transform analysis, and calculate the gradient mean using the Sobel operator;

[0014] S2, the quality grading module is used to generate a quality label for the image, to calculate the voting average grade and the weighted average grade for the output data of the three filters, and to generate a quality label, wherein the output results of at least two filters indicate that the image quality is qualified, and the input image and the corresponding quality label are stored together in the image storage system;

[0015] S3, the visual transformer encoder combined with the fully connected network is trained to directly predict the quality score from the image and learn the linear prediction mapping from the image features to the quality score, and the trained model is applied in the inference pipeline to predict the quality score of a new image and automatically save the images meeting the quality standard for downstream applications.

[0016] As a further improvement of the present application, the Laplacian variance filter uses a predefined threshold that is adjustable based on application requirements, the FFT variance filter applies amplitude spectrum analysis in the frequency domain by shifting the zero-frequency component, computing the amplitude spectrum, and converting to a logarithmic scale, the gradient mean filter uses a Sobel operator to compute pixel-level gradient magnitudes in the x and y directions.

[0017] As a further improvement of the present application, the voting average rank requires a unanimous agreement between all three filter outputs before classifying the image as meeting the quality criteria.

[0018] As a further improvement of the present application, the visual transformer encoder processes the input image based on an attention-based mechanism that captures spatial relationships and visual patterns relevant to quality assessment.

[0019] As a further improvement of the present application, the image quality assessment and data generation method further comprises a logical operation that combines the outputs of the three filtering operations, operates through sequential filter application stages to generate a comprehensive quality assessment for grayscale input images, the grayscale images are directly applied to the Laplacian variance filter, which detects rapid intensity changes through second-derivative calculations on the image surface.

[0020] As a further improvement of the present application, the voting average rank calculation combines the outputs of the three filtering operations through a logical operation to determine a consensus-based quality measure. The voting average rank calculator uses a logical AND operation to combine the outputs of the multi-filter evaluation modules, generating a binary quality assessment indicating whether all three filtering methods agree on the image quality acceptability.

[0021] As a further improvement of the present application, it further comprises an automatic dataset generation system, the automatic dataset generation system comprises an image input system, the image input system is an interface for receiving images from various sources, including camera systems, file repositories or real-time video streams, and is adapted to different image formats, resolutions and color spaces, providing a consistent data representation for subsequent processing operations.

[0022] As a further improvement of the present application, the image input system also contains preprocessing capabilities to standardize the input images, including resolution scaling, format conversion and color space standardization, to ensure compatibility with downstream analysis components, the image input module is directly connected to the quality assessment component through a defined data path, enabling real-time processing of continuous image streams for real-time processing applications or batch processing scenarios for large image collections;

[0023] The automatic dataset generation system integrates a bidirectional communication architecture, including an image acquisition module, the image acquisition module includes a camera interface and an image preprocessing module;

[0024] As a further improvement of the application, the camera interface automatically modifies capture parameters, such as exposure settings, focus adjustment, or frame rate modification, according to the control signal to optimize image quality based on real-time assessment feedback.

[0025] As a further improvement of the application, the automatic dataset generation system comprises a quality assessment system, an image storage system, and a model training module.

[0026] As a further improvement of the application, the image storage system comprises a high-quality image repository and a quality label database, and the model training module comprises a feature extractor and a quality score predictor.

[0027] In summary, the cumulative dataset of images with corresponding filter values and quality labels is utilized to train the prediction models, which can learn the relationship between visual image features and quality measurements. Enabling the system to develop automatic quality assessment capabilities, replicate the assessment results of the multi-filter evaluation module without explicit filter calculations during subsequent inference operations, generate prediction algorithms, and produce continuous quality scores instead of binary classifications, achieve nuanced quality assessment in the spectrum of image features and application requirements by arranging interconnected module operations in a structured architecture for comprehensive image processing and quality assessment.

[0028] Through sequential filter application stage operations, the image storage system creates a comprehensive dataset that records the original filter measurements and derived quality assessments for each input image and the values of the individual Laplacian variance filter, FFT variance filter, and gradient mean filter, in addition to the VAG / WAG labels. The storage process preserves the complete analysis results of each filtering stage, enabling subsequent analysis of individual filter performance and providing detailed documentation for the quality assessment justification of each processed image. The storage of filter values along with quality labels creates a labeled dataset that can be trained as machine learning application data while maintaining traceability between input images and their corresponding quality measurements. The stored filter values are analyzed to refine threshold parameters or adjust the calculation method based on downstream application performance requirements.

[0029] The comprehensive bidirectional communication architecture of the automatic dataset generation system enables continuous operation with multidirectional data flow between components, creating a feedback loop that enhances system performance through iterative refinement and adaptive optimization procedures. The system coordinates image acquisition, quality assessment, dataset management, and model training operations through interconnected communication pathways, enabling real-time feedback mechanisms and dynamic parameter adjustments based on performance observations and application requirements. This bidirectional architecture facilitates continuous learning and adaptation, where each component influences upstream and downstream operations through feedback mechanisms, optimizing overall system performance across different imaging applications and operating conditions while maintaining consistent quality standards and processing efficiency throughout the complete workflow from image acquisition to model deployment and application utilization. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 Main structure diagram of the image quality evaluation system of the first embodiment of the present application;

[0031] Figure 2 Image quality evaluation diagram of the multi-filter image processing system of the first embodiment of the present application through a parallel analysis path;

[0032] Figure 3 Three-stage workflow diagram of the first embodiment of the present application;

[0033] Figure 4 Image quality evaluation and data set generation work interaction flowchart of the first embodiment of the present application;

[0034] Figure 5 Data flow diagram of the image quality evaluation system of the first embodiment of the present application;

[0035] Figure 6 Processing architecture diagram of the image quality evaluation system of the first embodiment of the present application;

[0036] Figure 7 Image quality evaluation diagram of the multi-filter image processing system of the second embodiment of the present application through a sequential analysis path;

[0037] Figure 8 Image quality evaluation and data set generation work interaction flowchart of the second embodiment of the present application;

[0038] Figure 9 Diagram of the automatic data set generation system of the third embodiment of the present application;

[0039] Figure 10 Bidirectional communication architecture diagram of the automatic data set generation system of the third embodiment of the present application. DETAILED DESCRIPTION

[0040] The three embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0041] First embodiment:

[0042] Figure 1 The multi-filter and deep learning-based image quality evaluation system shown includes an image input system, a multi-filter evaluation module, a quality grading module, a deep learning model, an image storage system, and an inference pipeline.

[0043] The multi-filter evaluation module includes a Laplacian variance filter, an FFT variance filter, and a gradient mean filter, the Laplacian variance filter detects rapid intensity changes in the grayscale image by computing variance and comparing with a predefined threshold, which is selected by a person skilled in the art according to the application scenario and precision requirement of the image quality evaluation system, the FFT variance filter applies 2D fast Fourier transform to the grayscale image and computes variance for threshold comparison; the gradient mean filter computes gradient using Sobel operator and determines the average gradient value for threshold comparison;

[0044] The quality grading module includes a voting average grade (VAG) calculator and a weighted average grade (WAG) calculator, which perform voting average grade and weighted average grade calculation for the three filter output data and generate quality labels, wherein the output results of at least two filters indicate that the image quality is qualified;

[0045] The deep learning model includes a visual transformer encoder and a fully connected network connected to each other, which establishes a neural network architecture for analyzing image features for quality prediction purposes, the visual transformer encoder processes the input image through an attention-based mechanism to capture spatial relationships and visual patterns related to quality evaluation, while the fully connected network maps the extracted features to a quality score corresponding to the VAG or WAG method established in the training phase, the deep learning model learns the linear prediction mapping from image features to quality scores through a supervised learning procedure that analyzes the relationship between visual image features and the quality labels generated by the corresponding filters of the quality grading module. This learning process enables the deep learning model to replicate the multi-filter evaluation results without the need for explicit calculation of individual filter values, thereby eliminating the need for manual filter calculation during inference operations. The hierarchical positioning of the deep learning model allows the system to switch from explicit analysis processing to learning pattern recognition, reducing computational overhead while maintaining quality evaluation accuracy under different imaging conditions and content types;

[0046] An image storage system stores each input image, corresponding filter output data, and quality labels;

[0047] The inference pipeline includes a quality score predictor and an image selector. The quality score predictor processes image features through a trained neural network architecture to generate a numerical quality assessment corresponding to the VAG or WAG scoring method. The quality score predictor produces continuous quality scores rather than binary classifications, enabling nuanced quality assessment in the context of image features and application requirements. The image selector automatically saves images that meet quality standards for downstream applications such as object detection, medical diagnosis, surveillance analysis, or other computer vision tasks that benefit from high-quality input data. The quality standards are determined by those skilled in the art based on the use field and use requirements. The image selector discards images that fail to meet established threshold requirements, which are also set by those skilled in the art based on the use scenario and use requirements of the image quality assessment system. This prevents low-quality data from entering the subsequent processing pipeline and saves storage resources and computing power for acceptable images.

[0048] Figure 2 The multi-filter image processing system is shown to employ a comprehensive approach to image quality assessment through parallel analysis paths:

[0049] After the image input system inputs an image frame, the frame is converted to grayscale format to eliminate color variations that introduce inconsistencies in quality measurement, allowing the system to focus on structural and textural features that indicate image sharpness and acuity. The conversion process transforms color images into single-channel representations while preserving luminance information relevant to quality assessment purposes, establishing a standardized foundation for subsequent filtering operations.

[0050] After grayscale conversion, the system branches into three distinct parallel analysis paths, each designed to capture different aspects of image quality through specialized filtering techniques:

[0051] The first analysis path applies a Laplacian operation to the grayscale image, detecting rapid intensity changes and edge information by calculating second-order derivatives across the image surface. This Laplacian analysis provides a measure of local intensity variations related to image sharpness and focus quality.

[0052] The second analysis path performs a Fast Fourier Transform (FFT) analysis on the grayscale image, converting spatial domain information into frequency domain representations that reveal spectral features related to image clarity and noise content.

[0053] The third analysis path conducts gradient analysis, examining directional intensity variations across the entire image to assess edge strength and overall structural definition.

[0054] The outputs of these three parallel analysis paths are combined through two different averaging methods, providing complementary quality assessments. The Laplacian and FFT outputs are input into a voting average calculator, applying a consensus-based logic operation to determine a quality measurement based on agreement between the two filtering methods, while the FFT and gradient outputs are input into a weighted average calculator, which uses a summation-based standard to combine these measurements. This dual averaging approach allows the system to generate multiple quality indicators from the same input image, providing redundancy and enhanced reliability in quality determination. The combination of different filtering techniques with various averaging methods creates a robust evaluation framework that can capture quality characteristics overlooked by single-filter methods, resulting in a more comprehensive and accurate image quality assessment under different imaging conditions and content types.

[0055] Figures 3-6 An automated image quality assessment and dataset generation three-stage workflow is shown, in which images are evaluated through a multi-filter evaluation module, generating quantitative measurements characterizing various aspects of image quality, including sharpness, frequency content, and gradient information. The multi-filter evaluation stage produces data from each filtering operation as input parameters for subsequent quality assessment procedures, a voting average rank calculator component receives filter outputs from all three parallel filtering components and applies a consensus-based logic and operation to determine a quality measurement based on agreement between filtering methods, requiring unanimous agreement between filtering methods before indicating an acceptable image quality level. A weighted average rank calculator component processes the same filter outputs through a summation-based standard, allowing for more flexible quality determination when a majority of filtering methods indicate acceptable quality characteristics. These dual computation paths provide complementary quality assessment methods that can capture different aspects of image acceptability based on different application requirements and quality tolerance levels.

[0056] The multi-filter evaluation module output values are evaluated against predefined quality criteria, in some cases involving threshold comparisons of individual filter outputs or their combinations, establishing minimum acceptable values for image clarity and structure definitions. When an image satisfies the quality criteria, the image, filter values, and corresponding quality labels are stored to an image storage system, which records specific filter values and evaluation results, the entire storage process is automated based on multi-filter evaluation module analysis results to create quality label data with minimal human intervention, the image storage system enables the accumulation of processed image data with associated quality metadata, facilitating subsequent model training operations that can learn relationships between visual image features and multi-filter quality evaluation results, while the stored data can also be analyzed to refine threshold parameters or adjust aggregation methods based on downstream application performance requirements, allowing the system to adapt to different quality standards in different imaging domains and application contexts.

[0057] After which the deep learning model is re-initiated for model training, the system utilizes the accumulated dataset of images with corresponding filter values and quality labels to train prediction models that can learn the relationship between visual image features and quality measurements. Enabling the system to develop automated quality assessment capabilities, replicating the evaluation results of the multi-filter evaluation module without the need for explicit filter calculations during subsequent inference operations, generating prediction algorithms that can provide continuous quality scores instead of binary classifications, allowing for nuanced quality assessments across the image feature spectrum.

[0058] After model training is complete, the system applies the trained model to predict the quality score of incoming images, generating numerical quality assessments corresponding to the previously established VAG or WAG scoring methods, the inference pipeline produces quality predictions related to the multi-filter evaluation results without the need for explicit Laplacian variance, FFT variance, or gradient mean calculations.

[0059] The predicted quality score is compared against a predetermined threshold to determine the image's acceptability for downstream applications: when the predicted quality score meets or exceeds the established threshold, the image is stored for downstream applications such as object detection, medical diagnosis, or surveillance analysis. Conversely, when the predicted quality score falls below the threshold requirement, the corresponding image is discarded, preventing low-quality data from entering the subsequent processing pipeline. This threshold-based selection mechanism allows the system to maintain consistent quality standards while automatically filtering image datasets for various computer vision applications.

[0060] Second embodiment:

[0061] Figures 7-8 The multi-filter and deep learning-based image quality assessment system is shown to combine the outputs of the three filtering operations through a logical operation, generating a comprehensive quality assessment for the grayscale input image through sequential filter application stages. The grayscale image is directly applied to the Laplacian variance filter, which detects rapid intensity changes through second-order derivative calculations across the image surface. The Laplacian variance filter calculates variance measurements by analyzing local intensity variations related to image sharpness features, generating numerical values as quantitative indicators of edge definition and focus quality. This Laplacian variance calculation produces a threshold-comparable measurement indicating whether the image contains sufficient structural detail for downstream processing applications;

[0062] After Laplacian variance calculation, the FFT variance filter converts the grayscale image from a spatial domain representation to a frequency domain analysis, applying a FFT variance filter operation to perform a 2D Fast Fourier Transform on the grayscale image data, followed by a zero-frequency component shift to center the frequency spectrum for enhanced analysis precision. The FFT variance filter shifts the zero-frequency component to the center of the frequency spectrum prior to calculating the magnitude spectrum, allowing for more precise frequency domain measurements that reveal spectral features related to image sharpness and noise content, the FFT variance filter converts the magnitude spectrum to a logarithmic scale prior to calculating the variance, providing an enhanced dynamic range representation that can capture subtle frequency variations indicative of image quality differences;

[0063] Finally, the gradient mean filter uses a Sobel operator to calculate x and y gradients, measuring directional derivatives that convert image surface intensity in horizontal and vertical directions. The gradient mean filter calculates an amplitude value at each pixel location by combining the x and y gradient components, then determines an average gradient value for the entire image for threshold comparison purposes. This gradient analysis provides a measure of edge strength and overall structure definition that complements the frequency domain and second derivative information obtained from previous filtering stages.

[0064] The voting average grade calculation combines the outputs of the three filtering operations through a logical operation to determine a consensus-based quality measure. The voting average grade calculator uses a logical AND operation to combine the outputs of the multi-filter evaluation modules, generating a binary quality assessment that indicates whether all three filtering methods agree on image quality acceptability. This logical AND approach provides a conservative quality determination by requiring unanimous agreement between filtering methods before classifying an image as meeting quality standards, where the weighted average grade calculation applies a summation-based criterion to the filter outputs to obtain an alternative quality assessment method. The weighted average grade calculator requires the sum of the multi-filter evaluation module outputs to be greater than or equal to 2 to accept an image when a majority of filtering methods indicate an acceptable quality level.

[0065] The subsequent deep learning model and downstream application preparation are the same as the first embodiment, and will not be described again.

[0066] The image storage system creates a comprehensive dataset of raw filtering measurements and derived quality assessments for each input image and individual Laplacian variance filter, FFT variance filter, and gradient mean filter values, excluding the VAG / WAG labels. The storage process preserves the complete analysis results of each filtering stage, enabling subsequent analysis of individual filter performance and providing detailed documentation for the quality assessment justification of each processed image. The storage of filter values along with quality labels creates a labeled dataset that can serve as training data for machine learning applications while maintaining traceability between input images and their corresponding quality measurements. The stored filter values are analyzed to refine threshold parameters or adjust the calculation method based on downstream application performance requirements.

[0067] Third implementation:

[0068] Figure 9 The multi-filter and deep learning-based image quality assessment system is shown, further comprising an automatic dataset generation system, which includes an image input system as an interface for receiving images from various sources, including camera systems, file repositories, or real-time video streams, accommodating different image formats, resolutions, and color spaces, providing a consistent data representation for subsequent processing operations.

[0069] The image input system also contains preprocessing capabilities to standardize the input images, including resolution scaling, format conversion, and color space normalization, to ensure compatibility with downstream analysis components. The image input module is directly connected to the quality assessment component through defined data paths, enabling real-time processing of continuous image streams for live applications or batch processing scenarios for large image collections.

[0070] Referring to Figure 10 , a bidirectional communication architecture is shown that integrates the automatic dataset generation system, including an image acquisition module that serves as the primary interface for capturing and preprocessing images from various sources while maintaining a bidirectional communication path with downstream processing components. The image acquisition module coordinates image capture operations through specialized subcomponents that handle different aspects of data acquisition and initial processing procedures.

[0071] The image acquisition module includes a camera interface and an image preprocessing module, with the camera interface establishing a direct connection with imaging hardware, including digital cameras, video capture devices, or streaming video sources that provide continuous image data for quality assessment and dataset generation purposes. The camera interface manages hardware communication protocols, image format specifications, and capture timing parameters while providing a feedback mechanism for adjusting acquisition settings based on quality assessment results from downstream components.

[0072] The camera interface can receive control signals to modify capture parameters such as exposure settings, focus adjustments, or frame rate modifications to optimize image quality based on real-time assessment feedback.

[0073] The image preprocessing module applies initial conversion and normalization procedures to captured images before transmission to the quality assessment component through a bidirectional data path. The image preprocessing module performs operations such as format conversion, resolution scaling, noise reduction, or color space normalization while maintaining a communication channel with the quality assessment component that provides feedback on the effectiveness of preprocessing or optimization requirements for parameters. This bidirectional communication enables the image preprocessing module to adapt preprocessing algorithms based on quality assessment results, creating a feedback loop that enhances overall system performance through iterative refinement of image preparation procedures.

[0074] The image pre-processing module can also receive quality indicators from downstream components, influencing the selection of pre-processing parameters, and enabling dynamic optimization of the image preparation procedure based on observed quality assessment patterns and downstream application requirements.

[0075] The post-image multi-filter evaluation module, quality grading, and downstream application usage can be selected between the first and second embodiments.

[0076] The automated data set generation system includes a quality assessment system, an image storage system, and a model training module.

[0077] The image storage system includes a high-quality image repository and a quality label database. The high-quality image repository receives feedback from the model training system regarding image quality effectiveness or data set composition suggestions, enabling dynamic data set planning that adapts to evolving training requirements and application-specific quality standards. The bidirectional communication capabilities of the high-quality image repository enable continuous data set refinement through a feedback loop that optimizes storage procedures and data organization based on downstream utilization patterns and training effectiveness observations. The quality label database maintains a comprehensive record of quality assessment results and establishes bidirectional communication pathways with the quality assessment component and the model training system. The quality label database stores individual filter values, aggregated quality scores, and related metadata for each processed image, while providing query capabilities for selective data retrieval based on training requirements or analysis objectives. The quality label database maintains updated communication channels with the quality assessment system and supervisory learning data provision and training effectiveness feedback with the model training component. This bidirectional data exchange enables the quality label database to adapt the labeling procedures and metadata organization based on training performance observations and quality assessment refinement requirements.

[0078] The model training module includes a feature extractor and a quality score predictor, which handle feature extraction and quality prediction model development while providing feedback mechanisms to the quality assessment system regarding training data effectiveness and quality requirements. Within the model training module, the feature extractor processes stored images from the high-quality image repository, identifying visual patterns and features related to quality assessments saved in the quality label database, while maintaining communication channels with the data storage system for training data requests and quality feedback provision.

[0079] The integrated bidirectional communication architecture of the automated data set generation system enables continuous operation with multidirectional data flow between components, creating a feedback loop that enhances system performance through iterative refinement and adaptive optimization procedures. The system coordinates image acquisition, quality assessment, data set management, and model training operations through interconnected communication pathways, enabling real-time feedback mechanisms and dynamic parameter adjustments based on performance observations and application requirements. This bidirectional architecture facilitates continuous learning and adaptation, with each component influencing upstream and downstream operations through feedback mechanisms, optimizing overall system performance across different imaging applications and operating conditions while maintaining consistent quality standards and processing efficiency throughout the complete workflow from image acquisition to model deployment and application utilization.

[0080] In combination with the actual needs, the above-mentioned embodiments adopted by the present application are not limited to this, various changes made within the knowledge range of those skilled in the art without departing from the concept of the present application still fall within the protection scope of the present application.

Claims

1. A multi-filter and deep learning based image quality assessment system, characterized in that: The image quality assessment and data generation method comprises the following steps: S1, input the image to be evaluated through the image input system, and apply the image to the multi-filter evaluation module after gray processing, and calculate the Laplacian variance by detecting the rapid intensity change in the gray image, calculate the FFT variance by applying the 2D fast Fourier transform analysis, and calculate the gradient mean value using the Sobel operator; S2, use the quality grading module to generate a quality label for the image, calculate the voting average grade and weighted average grade for the output data of the three filters, and generate a quality label, wherein the output results of at least two filters indicate that the image quality is qualified, and the input image and the corresponding quality label are stored together in the image storage system; S3, train the visual transformer encoder combined with the fully connected network to directly predict the quality score from the image and learn the linear prediction mapping from the image features to the quality score, and apply the trained model in the inference pipeline to predict the quality score of the new image and automatically save the images meeting the quality standard for downstream applications. The Laplacian variance filter uses a predefined threshold that can be adjusted based on application requirements, the FFT variance filter applies amplitude spectrum analysis in the frequency domain by shifting the zero frequency component, calculating the amplitude spectrum and converting to logarithmic scale, and the gradient mean filter calculates the pixel-level gradient amplitude in x and y directions using the Sobel operator. 2.The multi-filter and deep learning based image quality assessment system according to claim 1, wherein: The voting average grade requires unanimous agreement among the outputs of all three filters before classifying the image as meeting the quality standard. The visual transformer encoder processes the input image based on an attention-based mechanism that captures spatial relationships and visual patterns related to quality assessment. The image quality assessment and data generation method further comprises a logical operation that combines the outputs of the three filtering operations, generates a comprehensive quality assessment for the gray input image through sequential filter application stages, and applies the gray image directly to the Laplacian variance filter to detect rapid intensity changes through second-order derivative calculation on the image surface. The automatic dataset generation system comprises an image input system and an image acquisition module, the image input system is an interface for receiving images from various sources, and the image input system further comprises preprocessing capabilities, the image acquisition module comprises a camera interface and an image preprocessing module. 3.The multi-filter and deep learning based image quality assessment system according to claim 2, wherein: The preprocessing capabilities of the image input system include resolution scaling, format conversion and color space standardization, and the image input module is signal connected with a quality assessment component. 4.The multi-filter and deep learning based image quality assessment system according to claim 2, wherein: The automatic dataset generation system comprises a quality assessment system, an image storage system and a model training module.

5. The multi-filter and deep learning based image quality assessment system according to claim 2, wherein: ​ 6.The multi-filter and deep learning based image quality assessment system according to claim 1, wherein: ​ 7. The multi-filter and deep learning based image quality assessment system according to claim 1, wherein: ​ 8. The multi-filter and deep learning based image quality assessment system according to claim 7, wherein: ​ 9.The multi-filter and deep learning based image quality assessment system according to claim 7, wherein: ​ 10. The multi-filter and deep learning based image quality assessment system according to claim 7, wherein: The image storage system includes a high-quality image storage repository and a quality label database, and the model training module includes a feature extractor and a quality score predictor.

Images