Wire rod steel surface defect detection system based on image denoising

Abstract

The invention relates to the technical field of wire rod steel detection, and discloses a wire rod steel surface defect detection system based on image denoising. The system comprises an image acquisition module, a fluctuation evaluation module, a noise probability calculation module, a frequency extraction module and an error derivation module. The image acquisition module acquires surface image data of the detection area; the fluctuation evaluation module determines a texture fluctuation index by comparing the gray statistical characteristics of the detection area and the adjacent area in the corresponding local windows; the noise probability calculation module evaluates the noise pollution possibility of each frequency component in combination with the frequency domain energy distribution difference, the frequency spectrum phase deviation and the texture fluctuation index; the frequency extraction module extracts a significant noise frequency accordingly; and the error derivation module calculates the defect detection error probability based on the aggregation characteristics of the significant noise frequencies of the adjacent regions. The system processes noise through cooperation of multiple modules, improves the detection reliability, and is suitable for wire rod steel quality detection in an industrial scene.

Description

Technical Field

[0001] This invention relates to the field of coiled steel inspection technology, specifically to a surface defect detection system for coiled steel based on image denoising. Background Technology

[0002] Coiled steel, as a fundamental raw material in industrial production, is widely used in construction, machinery manufacturing, automotive industry, rail transportation, and many other fields. Its surface quality directly affects the performance and safety of subsequent processed products. During the production of coiled steel, factors such as rolling process, cooling environment, and friction from transmission equipment can easily cause defects on the surface, such as cracks, scratches, oxide scale peeling, and inclusions. If these defects are not detected and addressed in a timely manner, they may lead to product scrap, equipment failure, or even safety accidents in subsequent processing stages. Therefore, accurate detection of surface defects in coiled steel is a crucial aspect of industrial quality control.

[0003] With the development of industrial automation technology, traditional inspection methods relying on manual visual inspection are gradually being replaced by automated inspection technologies. Manual inspection is not only affected by factors such as the subjective experience and visual fatigue of inspectors, resulting in low inspection efficiency and high rates of missed or false positives, but it also struggles to meet the high-speed inspection demands of modern production lines. Currently, among automated inspection technologies, machine vision-based surface defect detection methods have become the mainstream direction for surface defect detection of coiled steel due to their advantages such as non-contact operation, high speed, and quantifiability. These methods typically acquire images of the steel surface using image acquisition equipment, and then analyze the images using image processing algorithms to identify surface defects.

[0004] However, in real-world industrial scenarios, image acquisition is susceptible to various noise interferences, resulting in noise contamination in the acquired images of the coiled steel surface. For example, fluctuations in light levels in industrial environments can cause uneven image grayscale, creating illumination noise; heating of electronic components and signal transmission interference in image acquisition equipment can introduce electronic noise; and oxide scale reflections and oil adhesion on the coiled steel surface can generate texture noise. These noises can mask the true defect characteristics of the steel surface, making it difficult for subsequent defect identification algorithms to accurately distinguish between noise and defects, thus affecting the accuracy of the detection results.

[0005] To reduce the impact of noise on detection, existing technologies often employ image denoising algorithms to preprocess acquired images. Common denoising methods include spatial domain-based mean filtering and median filtering, as well as frequency domain-based low-pass filtering and wavelet transform denoising. However, these traditional denoising methods have significant limitations: spatial domain-based filtering methods often rely on the statistical characteristics of local pixels for noise suppression, easily blurring the edges of subtle defects on the steel surface, leading to the loss of defect information; frequency domain-based filtering methods typically preset fixed filtering thresholds or frequency ranges, making it difficult to dynamically adjust according to the noise characteristics under different detection scenarios. When the noise frequency is close to the defect characteristic frequency, incomplete noise removal or misinterpreting defect signals as noise can easily occur.

[0006] Existing inspection systems have shortcomings in noise assessment and error control. Most systems only process noise in images of a single inspection area, failing to consider the correlation of image features between adjacent inspection areas. This results in a lack of holistic noise assessment and difficulty in accurately evaluating the distribution pattern of noise in the image. Furthermore, existing systems rarely directly correlate noise characteristics with defect detection errors, failing to quantify the reliability of inspection results based on the actual impact of noise. This makes it difficult for inspectors to determine the confidence level of the results, increasing the difficulty of subsequent quality assessment. These problems collectively limit the performance of existing machine vision-based surface defect inspection systems for coiled steel in complex noisy environments, making it difficult to meet the high-precision and high-reliability industrial inspection requirements. Summary of the Invention

[0007] The purpose of this invention is to provide an image denoising-based surface defect detection system for coiled steel to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides a surface defect detection system for coiled steel based on image denoising, the system comprising: The image acquisition module is used to acquire surface image data of the inspection area of ​​the coiled steel. The fluctuation evaluation module is used to determine the texture fluctuation index of the detection area in each local window based on the gray-scale statistical characteristics of the surface image data of the detection area in the local window and the gray-scale statistical characteristics of the surface image data of adjacent detection areas in the corresponding local windows. The noise probability calculation module is used to combine the energy distribution differences, spectral phase shifts, and texture fluctuation indices of the surface image data of all local windows in adjacent detection areas in the frequency domain to evaluate the possibility of noise contamination in the surface image data of the detection area at each frequency component. The frequency extraction module is used to extract significant noise frequencies based on the noise contamination probability of all frequency components of the surface image data of the detection area. The error derivation module is used to calculate the probability of defect detection error based on the clustering characteristics of significant noise frequencies in all adjacent detection areas.

[0009] Preferably, the fluctuation assessment module determines the texture fluctuation index by: dividing the surface image data into multiple local windows; for each local window, calculating the overlap area of ​​the gray-level histogram of the detection region and its adjacent detection regions within the window; and using the normalized value of the overlap area as the texture fluctuation index of the detection region in that local window.

[0010] Preferably, the method for dividing the local window is as follows: the surface image data is decomposed into pyramid-style multi-scale decomposition to obtain image layers of different resolutions, and each image layer serves as a local window.

[0011] Preferably, the noise probability calculation module assesses the likelihood of noise contamination by: for any two adjacent detection areas, using each local window of the surface image data of the preceding detection area as a reference window; identifying the set of pixels in the surface image data of the following detection area that are spatially aligned with the reference window to form a comparison window; calculating the energy ratio and phase angle difference between the reference window and the comparison window in the same frequency band, and deriving the frequency stability coefficient based on the energy ratio and phase angle difference; standardizing the quotient of the texture fluctuation index and the frequency stability coefficient, subtracting it from a fixed constant, and averaging it over all local windows to obtain the noise contamination likelihood of each frequency component.

[0012] Preferably, the frequency stability coefficient is derived by multiplying the energy ratio and the phase angle difference by weighting factors and then adding them together. The sum is used as the frequency stability coefficient, wherein the frequency stability coefficient is directly proportional to both the energy ratio and the phase angle difference.

[0013] Preferably, the frequency extraction module extracts significant noise frequencies by: calculating the noise contamination probability of each frequency component within all local windows of the surface image data of each detection area, and marking frequency components with a noise contamination probability exceeding a preset threshold as significant noise frequencies.

[0014] Preferably, the error derivation module calculates the defect detection error probability by: collecting all significant noise frequencies of the surface image data of each detection area; obtaining a common subset of the significant noise frequencies of all detection areas, and using the frequencies in the common subset as key frequencies; calculating the dispersion of the amplitude values ​​of the surface image data of each detection area and its adjacent detection areas at the key frequencies, and taking the average of the dispersion as the noise-induced error of the detection area; normalizing the noise-induced errors of all detection areas, and using the normalized result as the defect detection error probability.

[0015] Preferably, the preset threshold is an adaptive threshold, and its determination method includes: For a specific frequency component within all local windows of a detection area, calculate the average and standard deviation of all noise contamination probabilities. The average value plus N times the standard deviation is used as the preset threshold for the frequency component, where N is a configurable constant greater than zero.

[0016] Preferably, the pyramid-style multi-scale decomposition employs the Gaussian pyramid algorithm, specifically including: continuously performing Gaussian smoothing and downsampling operations on the original surface image data to generate image representations at multiple scales, with each scale image representation serving as an analysis level.

[0017] Preferably, the calculation of the overlap area of ​​the gray-level histograms of the detection area and its adjacent detection areas within the window includes: quantizing the gray-level values ​​within the local window into 256 levels; calculating the gray-level histograms of the detection area and the adjacent detection areas respectively, and normalizing the histograms so that the sum is 1; for each gray-level, taking the minimum value of the two histogram values; and summing the minimum values ​​of all gray-levels to obtain the overlap area.

[0018] Compared with the prior art, the beneficial effects of the present invention are: From the perspective of image data processing, the image acquisition module can reliably acquire surface image data of the inspection area of ​​the coiled steel, providing basic image information for the analysis and processing of all subsequent modules. Unlike traditional inspection systems that only focus on the acquisition of single image data, the image data acquired by this module can be directly connected to the subsequent fluctuation assessment module, ensuring the continuity of the data flow throughout the entire inspection process, avoiding processing interruptions caused by data format mismatch or missing information, and guaranteeing the continuity of the inspection process.

[0019] The fluctuation assessment module accurately captures the texture fluctuation of coiled steel surface images by introducing the grayscale statistical features of local windows and comparing the grayscale statistical features of the detected area with those of adjacent detection areas within the corresponding local windows. In coiled steel surface detection, texture fluctuations can be caused by differences in the steel's own surface texture or by noise interference. This module effectively distinguishes between texture changes caused by these two factors through dual comparison of local windows and adjacent areas. The determined texture fluctuation index directly reflects the image stability of different areas, providing targeted feature basis for subsequent noise probability calculation. This texture fluctuation analysis method based on adjacent area comparison overcomes the one-sidedness of local information caused by analyzing only a single area in traditional detection techniques, making the judgment of image texture features more holistic and accurate.

[0020] The noise probability calculation module combines frequency domain energy distribution differences, spectral phase shifts, and texture fluctuation indicators to assess the likelihood of noise contamination in various frequency components of the surface image data of the detection area from multiple dimensions. Traditional noise assessment methods often rely on single frequency domain or spatial domain features, making it difficult to comprehensively reflect the complex characteristics of noise. This module, however, combines frequency domain energy distribution with spectral phase information, capturing the energy variation patterns and phase anomalies of noise in the frequency dimension. It also incorporates texture fluctuation indicators to further correlate with spatial domain image features, forming a multi-dimensional noise assessment system. This multi-factor fusion assessment method can more accurately identify the probability of noise presence in different frequency components, avoiding noise omissions or over-assessments caused by misjudgments based on single features, and providing a reliable basis for subsequent extraction of significant noise frequencies.

[0021] The frequency extraction module extracts significant noise frequencies based on the noise contamination probability of each frequency component, accurately locating the frequency range corresponding to the main noise sources in the surface image of coiled steel. Traditional denoising techniques often use a fixed frequency range for filtering, which cannot be dynamically adjusted according to the actual noise distribution. However, the significant noise frequencies extracted by this module are directly based on the previous noise contamination probability assessment results. It can dynamically determine the noise frequencies that need to be prioritized for processing based on the noise characteristics under different detection scenarios, making subsequent noise processing more targeted and avoiding over-processing of non-noise frequency components. This protects the characteristic information of steel surface defects within the corresponding frequency range and reduces the loss of defect information.

[0022] The error derivation module calculates the defect detection error probability based on the clustering characteristics of significant noise frequencies in all adjacent detection areas, quantifying the impact of noise on defect detection results. Traditional detection systems often only provide defect detection results without assessing their reliability, making it difficult for inspectors to determine whether the results are affected by noise. This module analyzes the clustering characteristics of significant noise frequencies in adjacent detection areas to discover patterns in the spatial distribution of noise, thereby inferring the potential error this noise distribution might cause in defect detection. This allows inspectors to clearly understand the error probability of the current detection results, providing a reference for subsequent quality assessment.

[0023] From the overall inspection process perspective, the system's modules are logically interconnected. From image acquisition to error calculation, the output of each module serves as the input for the next, forming a complete closed loop for noise processing and inspection error assessment. In industrial scenarios, this closed-loop processing mechanism can adapt to the inspection needs of coiled steel under different working conditions. Whether it's noise interference caused by changes in lighting, equipment vibration, or differences in the steel surface condition, the system can effectively reduce the impact of noise on defect detection through multi-module collaboration, making the defect detection results more consistent with reality. Furthermore, the system does not rely on complex hardware upgrades; performance improvements can be achieved simply by optimizing the algorithm module. It is easy to promote and apply in existing industrial production lines, seamlessly integrating with existing production processes, reducing the technical transformation costs for enterprises, and improving production efficiency and quality control levels while ensuring the effectiveness of coiled steel surface defect detection. Attached Figure Description

[0024] Figure 1 This is a timing diagram of the image denoising-based surface defect detection system for coiled steel described in this invention. Figure 2 A flowchart for a noise pollution potential assessment; Figure 3 This is a flowchart for calculating the probability of defect detection errors. Detailed Implementation

[0025] 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.

[0026] Please see Figure 1 This invention provides a surface defect detection system for coiled steel based on image denoising. The system includes: an image acquisition module, a fluctuation assessment module, a noise probability calculation module, a frequency extraction module, and an error derivation module. Specific implementation details are as follows: The image acquisition module is responsible for acquiring surface image data of the inspection area of ​​the coiled steel. This data is typically captured by a high-resolution industrial camera under uniform lighting conditions to ensure consistent image quality. The fluctuation assessment module calculates a texture fluctuation index based on the gray-level statistical characteristics of the surface image data in a local window of the inspection area and the gray-level statistical characteristics of the surface image data in corresponding local windows of adjacent inspection areas. This index reflects the degree of local variation in surface texture. The noise probability calculation module combines the differences in energy distribution, spectral phase shift, and texture fluctuation index of the surface image data in all local windows of adjacent inspection areas in the frequency domain to assess the likelihood of noise contamination in each frequency component of the surface image data of the inspection area, thereby quantifying the impact of noise on the image. The frequency extraction module extracts significant noise frequencies based on the likelihood of noise contamination; these frequency components are identified as noise sources that may interfere with defect detection. The error derivation module calculates the defect detection error probability by analyzing the clustering characteristics of significant noise frequencies in all adjacent inspection areas, providing a reliability assessment for the final defect judgment.

[0027] Example 1: After the image acquisition module acquires the surface image data of the inspection area of ​​the coiled steel, the fluctuation assessment module begins to process this data. The fluctuation assessment module divides the surface image data into multiple local windows. The local windows are divided using a regular grid method. Each window covers a continuous area of ​​the surface image data. The window size is adjusted according to the image resolution to ensure effective capture of local features. The selection of the window size is based on prior knowledge of image resolution and defect size. Typically, the window size is set to 32x32 pixels or 64x64 pixels to balance computational efficiency and detail preservation. For each local window, the fluctuation assessment module calculates the overlap area of ​​the gray-level histogram of the inspection area and its adjacent inspection areas within the window. When calculating the overlap area, the gray-level values ​​within the local window are quantized into 256 levels. The quantization process is based on a linear mapping to convert the original gray-level value range to an integer space of 0-255. The linear mapping ensures a uniform distribution of gray-level values ​​and avoids quantization deviations affecting subsequent analysis. The gray-level histograms of the inspection area and adjacent inspection areas are calculated separately. The histogram calculation is performed by statistically analyzing the image of each gray-level within the window. The frequency of pixel occurrence is calculated using an efficient accumulation algorithm. It iterates through each pixel within the window, updating the frequency count of the corresponding level based on its grayscale value. After histogram calculation, normalization is performed to ensure the sum is 1. Normalization is achieved by dividing the frequency of each grayscale level by the total number of pixels within the window, ensuring consistent histogram proportions and facilitating cross-window comparisons. For each grayscale level, the minimum of two histogram values ​​is taken. This minimum value operation is implemented through element-wise comparison, which is optimized using parallel computing to accelerate processing. The minimum values ​​of all grayscale levels are summed to obtain the overlapping area, which ranges from 0 to 1. The closer the value is to 1, the more similar the textures of the two detection regions. The normalized overlapping area is used as the texture fluctuation index for that detection region within that local window. The normalization process uses linear scaling to map the overlapping area to the 0-1 range. The linear scaling formula is simple and easy to implement, ensuring the dimensionless nature of the index. The fluctuation evaluation module repeats the above process for all local windows of each detection region, generating a texture fluctuation index matrix, which is used as input for the subsequent noise analysis module.

[0028] The division of local windows needs to consider the spatial continuity of surface image data. The window size is selected based on prior knowledge of image resolution and defect size. A window size of 32x32 pixels or 64x64 pixels is set to balance computational efficiency and detail preservation. A regular grid is used for window division, with the grid spacing set according to the window size to ensure minimal or no overlap between windows to avoid redundant calculations. The calculation of grayscale histograms employs efficient statistical algorithms to avoid memory overflow issues, especially in high-resolution image processing. Histogram calculation is accelerated using lookup tables, with quantized grayscale values ​​used as indexes to directly access pre-calculated data structures. The calculation of overlapping areas relies on histogram alignment to ensure that the histograms of two detection regions are in the same grayscale range. In terms of grayscale level comparison, the grayscale quantization step is implemented through linear mapping. The mapping function maps the original grayscale values ​​to the range of 0-255, and the mapping process takes into account the dynamic range adjustment of the image. Normalization ensures the dimensionless nature of the texture fluctuation index, which facilitates cross-image comparison. Normalization is performed by dividing by the total number of pixels, and the total number of pixels is calculated based on the fixed window size. The output of the fluctuation evaluation module is a sequence of texture fluctuation indices, with each index corresponding to a local window. The sequence length is equal to the number of local windows. Low values ​​of the texture fluctuation index indicate significant local texture changes, which may be due to surface defects or noise interference, while high values ​​indicate good texture consistency. The processing results of the fluctuation evaluation module are passed to the noise probability calculation module as one of the inputs for noise evaluation.

[0029] The division of local windows and the calculation of overlapping areas are the core of fluctuation assessment. Their implementation relies on efficient functions from image processing libraries such as OpenCV to ensure real-time processing capabilities. OpenCV provides histogram calculation and normalization functions, which can be directly called to simplify the development process. The grayscale quantization process uses integer operations to improve speed, and the quantization level of 256 is selected based on a standard grayscale image format, compatible with most industrial camera outputs. Histogram statistics use multi-threading, dividing window data into blocks for parallel computation, leveraging the advantages of multi-core processors. The minimum summation operation in the overlapping area calculation is optimized using vectorized instructions, achieving high-speed execution on modern CPUs. The normalization of texture fluctuation indices uses scalar multiplication with a scaling factor of 1.0 divided by the maximum possible overlap area, which theoretically has a value of 1.0. The algorithm complexity of the fluctuation evaluation module is proportional to the number and size of windows. Optimization measures include reducing unnecessary memory copying and using cache-friendly data structures. The grid division of local windows allows for configurable parameters, allowing users to adjust the window size and grid density according to specific applications to adapt to different detection needs. The calculation of the grayscale histogram considers the image boundary situation. For windows that extend beyond the image boundary, a filling strategy is adopted, such as mirror filling or constant filling, to maintain data integrity.

[0030] The implementation of the fluctuation evaluation module emphasizes robustness and efficiency. Local window partitioning uses a fixed-step sliding window approach, with the step size set smaller than the window size to increase sampling density and improve the accuracy of texture change capture. A quantization level of 256 is the standard choice for the grayscale histogram, but the system supports configurable quantization levels, allowing users to adjust the number of levels according to image characteristics to balance accuracy and computational load. Histogram normalization uses double-precision floating-point numbers to ensure computational accuracy and avoid the accumulation of rounding errors. The minimum value operation in overlapping area calculation uses pixel-by-pixel comparison, and the comparison algorithm is optimized to a branch-reduction version to improve CPU pipeline efficiency. Texture wave... The generation process of dynamic indicators includes a data verification step, which checks the validity of the input data, such as whether the window is empty or the histogram is all zeros, to avoid division by zero errors or invalid outputs; the fluctuation evaluation module integrates an error handling mechanism, which logs when abnormal data is detected and continues processing with default values ​​to ensure system stability; the processing order of local windows follows the raster scan order, from the upper left corner of the image to the lower right corner, and sequential processing simplifies the data access mode and improves the cache hit rate; the calculation of the grayscale histogram uses histogram equalization technology to preprocess the image and enhance local contrast, but in this embodiment, the original grayscale values ​​are maintained to faithfully reflect the texture features.

[0031] The parameters of the fluctuation evaluation module are optimized based on a large amount of test data. The window size is determined through cross-validation to balance detection sensitivity and computational resources. The grayscale quantification linear mapping function uses a fixed formula, and the mapping coefficients are automatically adjusted according to the image bit depth, supporting 8-bit, 12-bit, or 16-bit grayscale images. The frequency counting of histogram statistics uses atomic operations to avoid multi-threaded race conditions and ensure data consistency. The results of the overlap area calculation are stored using floating-point numbers, and the accuracy meets the requirements of subsequent modules. The normalization processing of the texture fluctuation index includes clipping operations to limit the index value to the range of 0-1 to prevent overflow. The output format of the fluctuation evaluation module is designed as a multi-dimensional array, with the array dimensions corresponding to the image spatial location and window index, which facilitates index access by subsequent modules. The division of local windows supports non-uniform grids, and the grid size is dynamically adjusted according to the importance of the image region. Smaller windows are used in important regions to improve resolution. The calculation of the grayscale histogram includes smoothing processing. Gaussian filtering is applied to the histogram to reduce the impact of noise, but smoothing is omitted in this embodiment to maintain the original statistical characteristics. The overlap area calculation uses approximate algorithms for acceleration, such as integral histogram technology, but this embodiment uses precise calculation to ensure accuracy.

[0032] The fluctuation assessment module is implemented in C++ using the OpenCV library. Key functions include window partitioning, histogram calculation, overlap area calculation, and normalization. The window partitioning function accepts image data and window size parameters, returning a list of window coordinates. The histogram calculation function accepts image patches and quantization level parameters, returning a normalized histogram vector. The overlap area calculation function accepts two histogram vectors, returning a scalar overlap area value. The normalization function accepts the overlap area value and returns a texture fluctuation index. The module interface is designed in an object-oriented style, encapsulating internal state and providing thread-safe methods. The performance of the fluctuation assessment module is analyzed. Analysis shows that most of the time is spent on histogram statistics and overlapping area calculation, and the optimization focus is on these hot spots; the division of local windows uses multi-scale expansion of the image pyramid concept, but this embodiment insists on single-scale grid division to simplify implementation; the calculation of grayscale histogram supports multiple color spaces, but this embodiment only processes grayscale images, and color images are first converted to grayscale space; the minimum value operation in the overlapping area calculation uses SIMD instructions to process multiple grayscale levels in parallel to improve throughput; the normalization processing of texture fluctuation index includes a standardization step to convert the index into a zero-mean unit variance distribution, but this embodiment uses simple linear normalization.

[0033] The fluctuation assessment module was tested using synthetic images and real images of coiled steel. Synthetic images simulated ideal textures and defect patterns to verify the algorithm's correctness; real images were collected from industrial environments, encompassing various surface conditions, to test the module's robustness. Local window partitioning parameters were optimized through grid search to select the optimal window size and grid step size. Gray-level histogram calculations were verified by comparing the output with a reference implementation to ensure consistency. The accuracy of overlapping area calculations was tested using image pairs with known overlap to measure calculation errors. Normalization of the texture fluctuation index was performed to check boundary conditions, such as fully consistent or fully random textures, ensuring the index values ​​met expectations. The integration test of the fluctuation assessment module was conducted in conjunction with the entire system to check data flow and interface compatibility. Local window partitioning considered computational resource limitations, with large images processed in blocks to avoid memory shortages. Gray-level histogram calculations used a memory pool to manage temporary data, reducing dynamic allocation overhead. The results of overlapping area calculations were cached for repeated detection areas to optimize performance. Texture fluctuation index generation included quality assessment, triggering recalculation when index values ​​were abnormal.

[0034] The fluctuation assessment module's extensible design supports a plug-in architecture, allowing for the replacement of window partitioning algorithms or histogram calculation methods. The local window partitioning algorithm includes overlapping and non-overlapping window options; this embodiment uses non-overlapping windows to simplify the logic. Gray-level histogram calculation supports weighted statistics, with pixel weights based on spatial location, but this embodiment uses uniform weights. The minimum summation operation in overlapping area calculation uses recursive decomposition to handle large levels, but 256 levels are directly processed in a loop. Normalization of the texture fluctuation index includes non-linear transformation options, such as logarithmic scaling, but this embodiment adheres to linear transformation. The fluctuation assessment module's documentation details the inputs, outputs, and side effects of each function for easy maintenance and debugging. Local window partitioning uses image segmentation techniques to predefine regions of interest, but this embodiment processes the entire detection region. Gray-level histogram calculation includes a background subtraction step to eliminate the influence of illumination variations, but this embodiment assumes uniform illumination. Overlapping area calculation uses advanced metrics such as mutual information, but this embodiment chooses simple overlapping area to maintain efficiency. Texture fluctuation index generation includes confidence calculation, with index values ​​accompanied by uncertainty estimation, but this embodiment outputs point estimates.

[0035] The deployment environment of the fluctuation assessment module considers the hardware configuration of industrial computers, and the code is optimized for x86 and ARM architecture compatibility; the division of local windows uses GPU acceleration, uploading image data to video memory for parallel processing, but this embodiment uses CPU implementation; the calculation of grayscale histograms uses the OpenCL framework for cross-platform execution, but this embodiment relies on OpenCV CPU functions; the minimum value operation in the overlapping area calculation uses multi-threaded block processing, dividing the level range into different threads; the normalization processing of texture fluctuation indicators uses lookup tables to accelerate floating-point operations, and pre-calculates the normalization mapping; the fluctuation assessment module... Block monitoring includes performance counters that record processing time and memory usage for system tuning; local window division supports dynamic adjustment, adapting the window size to the image content, but this embodiment uses a fixed size; grayscale histogram calculation includes histogram matching preprocessing, aligning the detection region histogram to the reference histogram, but this embodiment directly uses the original histogram; the result of overlapping area calculation is compressed to reduce memory usage, and the index matrix is ​​stored in a sparse format if most values ​​are zero; texture fluctuation index generation includes filtering postprocessing, applying a moving average to smooth the index sequence, but this embodiment outputs the original index.

[0036] Example 2: The local window partitioning method employs pyramid-style multi-scale decomposition. Pyramid-style multi-scale decomposition performs multi-resolution analysis on surface image data, generating image layers of different resolutions. Each image layer serves as a local window, providing texture feature representations from coarse to fine. The pyramid-style multi-scale decomposition uses the Gaussian pyramid algorithm, which specifically includes continuous Gaussian smoothing and downsampling operations on the original surface image data. Gaussian smoothing uses Gaussian convolution kernels to filter the image, with the kernel size and standard deviation adjusted according to the image noise level. The smoothing operation reduces the impact of high-frequency noise on the decomposition. The downsampling operation halves the image size through interval sampling, generating lower-resolution image layers. The Gaussian smoothing and downsampling process is repeated to generate image representations at multiple scales. Each scale image representation serves as an analysis level, and the number of analysis levels is determined by the original image size, typically 3 to 5 levels to cover features at different scales. Each analysis level serves as a local window for fluctuation assessment and noise analysis, capturing global and local changes in texture in a multi-scale manner.

[0037] The Gaussian pyramid algorithm prioritizes computational efficiency. Gaussian convolution employs separable convolution optimization to reduce computational load. Separable convolution decomposes a two-dimensional convolution into two one-dimensional convolutions, filtering sequentially in the row and column directions, significantly reducing computational complexity. Downsampling uses bilinear interpolation to maintain image smoothness. Bilinear interpolation calculates new pixel values ​​based on a weighted average of adjacent pixels, reducing aliasing effects introduced by downsampling. Pyramid-style multi-scale decomposition ensures scale invariance in local windows, enabling the handling of defect features of varying sizes. Scale invariance is achieved through multi-level analysis: large-scale levels capture macroscopic textures, while small-scale levels capture microscopic details. Image data at each analysis level is processed independently. The fluctuation assessment module calculates texture fluctuation indices at each level, using the same calculation method as in the previous embodiments, but adapting to image data of different resolutions. Multi-scale decomposition enhances the system's robustness to defects, as defects may manifest at multiple scales; this improved robustness stems from the multi-level representation of features. The parameter configuration of pyramid-style multi-scale decomposition affects the analysis results. The size of the Gaussian convolution kernel is chosen based on image resolution and noise characteristics. Larger kernels provide stronger smoothing but lose details, while smaller kernels preserve details but have weaker noise suppression capabilities. The standard deviation of the kernel controls the width of the Gaussian distribution; the larger the standard deviation, the smoother the distribution. Typically, the standard deviation is set to one-sixth of the kernel size to ensure energy concentration. The downsampling factor is fixed at 2, and the size is halved with each downsampling to maintain the regularity of the pyramid structure. The number of analysis levels is controlled by the iteration termination condition. Decomposition stops when the image size is smaller than a preset threshold, which is usually set to 16x16 pixels to ensure the effectiveness of the minimum window. The output of pyramid-style multi-scale decomposition is a set of multi-resolution images. These images are used as local windows input to the noise probability calculation module. The multi-resolution images are stored in a pyramid data structure for easy access across scales.

[0038] Multi-scale analysis improves the accuracy of noise assessment because the frequency domain representation of noise and defects varies with scale; large-scale levels mainly contain low-frequency components, while small-scale levels contain high-frequency components. Multi-scale partitioning of local windows is key to system adaptability, ensuring reliable operation under various imaging conditions. This adaptability is reflected in its ability to handle input images of different resolutions. The parameters of the Gaussian pyramid algorithm, such as the kernel size and downsampling factor, are configurable to adapt to different application scenarios; configuration information is input through a parameter file. The fluctuation assessment module requires adjusting the window size when operating on multi-scale local windows. While the physical size of the local window remains consistent across different image resolutions, the pixel size varies. The preprocessing steps of the pyramid-style multi-scale decomposition include image normalization, converting the input image to a standard brightness range to eliminate the influence of illumination variations. The extended functions of pyramid-style multi-scale decomposition include pyramid reconstruction, which progressively upsamples the image from a coarse-scale to restore the original resolution for subsequent image enhancement steps; the selection of Gaussian convolution kernels considers frequency domain characteristics, ensuring that the Gaussian kernel remains Gaussian in the frequency domain to avoid introducing ringing effects; the low-pass filtering characteristics of the downsampling operation prevent frequency aliasing, and the combination of anti-aliasing filtering and Gaussian smoothing ensures that the sampling theorem is satisfied; the scale difference between analysis levels is a power of 2, simplifying coordinate mapping calculations, and the scale factor is implemented using binary shift operations; the real-time requirements of pyramid-style multi-scale decomposition affect the selection of the number of levels, and industrial inspection typically uses a 3-level pyramid to balance latency and accuracy; the local windows are organized in a tree structure in the multi-scale hierarchy, with each window having a corresponding parent window at the coarse scale and child windows at the fine scale, forming a multi-resolution representation.

[0039] The numerical stability of the Gaussian pyramid algorithm is guaranteed by floating-point precision, and double precision is used for convolution calculation to avoid accumulated errors. The initialization steps for pyramid construction include boundary processing, which mirrors and fills the image boundaries to ensure consistent convolution result sizes. The parallelization of multi-scale decomposition is implemented using thread pool management, with each scale allocated an independent thread for parallel construction. The interface design of the pyramid-style multi-scale decomposition and fluctuation evaluation module is designed with a unified data format, converting all data to standard pixel blocks for processing regardless of scale. The sampling strategy for local windows in multi-scale analysis includes overlapping sampling, where adjacent windows partially overlap to increase sampling density. Variants of the Gaussian pyramid algorithm include the Laplacian pyramid, but this implementation uses only the Gaussian pyramid for simplicity. The quality assessment of pyramid-style multi-scale decomposition measures information loss by comparing the differences between the original image and the pyramid-reconstructed image through reconstruction error checks. Feature extraction of multi-scale local windows includes inter-scale difference calculation, and subtraction of images at adjacent scales yields a detail layer for defect enhancement. The discretization of Gaussian convolution kernels uses binomial coefficients for approximation, and integer kernel coefficients improve computational speed. Pixel selection for downsampling operations adopts a center-aligned method to maintain the geometric center of the image. The hardware acceleration scheme for pyramid-style multi-scale decomposition uses FPGA to implement the convolution and downsampling pipeline. The multi-scale integration strategy of local windows includes feature fusion, which weights and merges texture fluctuation indicators at different scales. The scale range of pyramid-style multi-scale decomposition can be dynamically adjusted, automatically selecting the optimal scale level based on the defect size.

[0040] The illumination invariance of the Gaussian pyramid algorithm is enhanced through local contrast normalization, with local brightness adjustments performed on the image at each scale. The storage format for pyramid construction adopts the multi-resolution TIFF standard, compatible with existing image processing libraries. The computational complexity of multi-scale analysis is proportional to the total number of pixels, and optimization measures include sparse sampling of unimportant regions. The texture descriptors of local windows at multiple scales include SIFT features, but this implementation uses grayscale statistical features to maintain consistency. The termination condition for pyramid-style multi-scale decomposition can be based on information entropy; decomposition stops when the entropy value of a new level changes less than a threshold. The Gaussian pyramid algorithm is implemented on GPUs using CUDA kernel functions, with parallel convolution and downsampling significantly improving speed. The anomaly handling mechanism of pyramid-style multi-scale decomposition includes scale consistency checks to ensure correct geometric transformations of images at each level; the coordinate system of the multi-scale local window is the original image coordinate system, facilitating defect localization and backtracking; the boundary processing of Gaussian convolution adopts symmetrical expansion to reduce the impact of boundary effects; the interpolation method for downsampling operations can be either nearest neighbor or bicubic interpolation, but this implementation uses bilinear interpolation to balance quality and speed; the open interface of pyramid-style multi-scale decomposition allows for replacement of the core algorithm, such as using wavelet pyramids instead of Gaussian pyramids; the multi-scale analysis results of the local window are stored as multi-dimensional feature vectors, with the vector dimension corresponding to the number of scales; the real-time display function of pyramid-style multi-scale decomposition supports monitoring of images at each level for debugging and parameter adjustment.

[0041] Example 3: See Figure 2 The noise probability calculation module assesses the likelihood of noise contamination. Its inputs come from the texture fluctuation index output by the fluctuation assessment module and the surface image data from the image acquisition module. The module processes any two adjacent detection regions, using each local window of the surface image data from the preceding detection region as a reference window. The selection of the reference window is based on the order of the detection regions, typically processed according to the image acquisition time or spatial location sequence. The reference window represents the baseline region for the current analysis. It then identifies the set of pixels in the surface image data of the following detection region that are spatially aligned with the reference window, forming a comparison window. Alignment is achieved through coordinate mapping, ensuring that the reference window and comparison window cover the same physical area. Coordinate mapping uses affine transformation or direct pixel coordinate transformation, considering possible geometric deformations in the image. The calculation... The energy ratio and phase angle difference between the reference window and the contrast window within the same frequency band are calculated. Frequency band division uses Fast Fourier Transform (FFT) to convert the window image to the frequency domain. A frequency band is defined as a set of continuous frequency intervals, with interval division based on a logarithmic scale to adapt to human visual perception. The number of frequency bands is typically set to 10 to 20. The energy ratio is calculated as the ratio of the sum of squares of the amplitudes of the reference and contrast windows within the frequency band. The sum of squares represents the energy distribution, and the energy ratio reflects amplitude variation. The phase angle difference is calculated as the average difference in phase angle between the reference and contrast windows within the frequency band. The phase angle is obtained through the argument of the Fourier coefficients, and the phase angle difference reflects phase consistency. Frequency stability coefficients are derived based on the energy ratio and phase angle difference. These coefficients quantify the stability of frequency components. The derivation of the frequency stability coefficients uses the following formula:

[0042] in: Represents the frequency stability coefficient. Indicates the energy ratio. Indicates the phase angle difference. and These are weighting factors, set according to application requirements. The weighting factor for energy ratio is usually greater than that for phase angle difference because energy changes are more sensitive to noise. The frequency stability coefficient is directly proportional to both energy ratio and phase angle difference; that is, the larger the energy ratio or the larger the phase angle difference, the higher the frequency stability coefficient, indicating that the frequency component is more unstable. The quotient of texture fluctuation index and frequency stability coefficient is standardized. The standardization uses the Z-score method to convert the quotient into a zero-mean unit variance distribution. The Z-score calculation uses the mean and standard deviation of the quotients of all local windows. The difference is taken with a fixed constant and then averaged over all local windows. The fixed constant is set to an empirical value, such as 1.0, to adjust for noise sensitivity. The difference operation reduces the influence of the basic noise level. The noise pollution probability of each frequency component is obtained; the higher the value, the greater the risk of noise pollution.

[0043] The implementation of the noise probability calculation module emphasizes computational accuracy and efficiency. Image registration techniques are used to identify the reference and comparison windows, ensuring accurate spatial alignment. Registration methods include feature point matching or optical flow, but this implementation uses simple coordinate mapping assuming a rigid image transformation. The sum of squared amplitudes in the energy ratio calculation uses floating-point arithmetic to avoid overflow; the amplitude value comes from the magnitude of the Fourier transform, which is implemented using a fast algorithm such as FFT. The phase angle difference calculation involves cyclic interpolation to avoid angle entanglement; the phase angle difference is obtained by calculating the angle between two phase vectors. The formula for the frequency stability coefficient includes weighting factors. and The noise assessment bias can be influenced by adjusting the weighting factor through training data. The optimization process uses gradient descent to minimize the noise detection error. Standardization makes the probability of noise contamination comparable across different images, and averaging smooths local variations. The output of the noise probability calculation module is a noise contamination probability map, with the map dimensions corresponding to the frequency components and detection regions. This map is used by the frequency extraction module.

[0044] Frequency band division is based on frequency domain analysis. The Fast Fourier Transform (FFT) transforms each local window image from the spatial domain to the frequency domain, with the transform size matching the window size, typically a power of 2 for optimized computation. Frequency band boundaries are set considering human visual characteristics: low-frequency bands have narrower coverage, while high-frequency bands have wider coverage, as the human eye is more sensitive to low frequencies. In energy ratio calculations, the energy values ​​of the reference and contrast windows are normalized to the same total, eliminating the influence of overall brightness differences. Phase angle difference calculations use the average phase difference, with weighted averaging by amplitude values, where high-amplitude frequency components contribute more. Weighting factors in the frequency stability coefficient formula are also considered. and satisfy The constraints ensure a consistent range of coefficients, but this implementation allows for independent settings for flexible adjustment; the Z-score method in the standardization process assumes a normal distribution of quotients, but in practice, a distribution test may be required; the selection of fixed constants is based on historical data statistics, and the value of the constant affects the baseline level of the possibility of noise pollution.

[0045] The noise probability calculation module's parameter configuration includes the number of frequency bands, weighting factor values, and fixed constants. These parameters are input through a configuration file and can be adjusted on-site. The processing order of the reference window and contrast window is optimized to a batch processing mode, processing multiple windows at once to improve throughput. The energy ratio calculation uses logarithmic scaling to convert the ratio to decibels, enhancing numerical stability. The angle unit in the phase angle difference calculation is unified to radians to avoid confusion between degrees and radians. The frequency stability coefficient formula is implemented using vectorized operations, processing multiple frequency bands simultaneously. Standardization processing includes outlier pruning to avoid extreme values ​​affecting distribution statistics. The average value calculation uses an arithmetic mean, considering weighted averaging to emphasize important windows. The parallelization of the noise probability calculation module uses multithreading, with each frequency band calculated independently. Spatial alignment of the reference window and contrast window considers image distortion, and the alignment algorithm includes perspective transformation to correct lens distortion. The energy value integration in the energy ratio calculation is over the frequency band, and the integration method uses the trapezoidal rule for approximation. The phase angle difference calculation uses the phase correlation method to improve accuracy, estimating the phase difference through a cross-correlation function. The weighting factor in the frequency stability coefficient formula... and Dynamic adjustment is implemented, with higher frequency bands potentially receiving greater weight based on changes in frequency band importance; standardization uses robust statistics such as median and absolute deviation to reduce the impact of outliers; fixed-constant subtraction can be replaced with proportional scaling, adapting to image content; the average calculation of the probability of noise contamination includes confidence interval estimation, but this implementation outputs point estimation.

[0046] Error sources in the noise probability calculation module include registration error, Fourier transform leakage, and quantization error. Error control is achieved through algorithm optimization and parameter calibration. The selection strategies for the reference and contrast windows include overlapping windows to increase sampling density. The amplitude sum of squares calculation in the energy ratio calculation uses double-precision accumulation to prevent rounding errors. The phase expansion algorithm in the phase angle difference calculation handles phase jumps to ensure continuity. The linear assumption in the frequency stability coefficient formula may not hold, but it simplifies implementation. The Z-score method for normalization fails when the distribution is skewed, requiring a backup scheme. The window weights for the average value calculation are based on the window area, with larger windows contributing more. Hardware acceleration utilizes GPUs for Fourier transforms and vector operations. Data layout optimization for the reference and contrast windows improves cache efficiency by enabling continuous memory access. The frequency band integral in the energy ratio calculation is accelerated using a pre-computed weight table. The phase angle difference calculation uses a lookup table to approximate an inverse trigonometric function. The weighting factors in the frequency stability coefficient formula... and Storage is kept in constant memory to reduce access latency; statistical calculations for normalization use parallel reduction algorithms; mean calculation uses a pipelined design for overlapping memory transfers and computations. Validation is achieved through synthetic image testing, which simulates ideal noise patterns; alignment accuracy of the reference and contrast windows is measured using calibrated images; accuracy of energy ratio calculations is verified by comparing with theoretical values; error analysis of phase angle difference calculations uses known phase difference signals; sensitivity testing of the frequency stability coefficient formula involves observing changes in output based on weighting factors; distribution testing of normalization uses a normal probability plot; robustness testing of mean calculations incorporates noise to observe stability. Integration and data exchange with other system modules uses a unified interface; index management of the reference and contrast windows uses hash tables for fast lookups; energy ratio calculation results are cached to avoid duplicate calculations; intermediate values ​​of phase angle difference calculations are stored in complex number format; evaluation of the frequency stability coefficient formula uses JIT compilation optimization; parameter learning for normalization uses online learning algorithms; and real-time updates of mean calculations use a sliding window.

[0047] Example 4: The frequency stability coefficient is derived by multiplying the energy ratio and phase angle difference by weighting factors and then summing the results. The sum is used as the frequency stability coefficient, which is directly proportional to both the energy ratio and the phase angle difference. The frequency extraction module extracts significant noise frequencies and calculates the noise contamination probability of each frequency component within all local windows of the surface image data for each detection region. The noise contamination probability calculation is as described in the previous examples, based on the frequency stability coefficient and texture fluctuation index. Frequency components with a noise contamination probability exceeding a preset threshold are marked as significant noise frequencies. The preset threshold is set to a fixed value or an adaptive value. The marking operation is achieved through threshold comparison, generating a binary frequency mask. Significant noise frequencies represent frequency components that may contain noise, and these frequencies are suppressed in subsequent denoising processing. The weighting factor of the frequency stability coefficient can be optimized through training data, and adjusting the weighting factor affects the bias of noise assessment. The frequency extraction module processes all detection regions and frequency components, and the extraction of significant noise frequencies is a prerequisite for frequency domain filtering. The selection of a preset threshold balances the precision and recall of noise detection; a higher threshold reduces false alarms but may miss weak noise. The output of the frequency extraction module is a list of significant noise frequencies, which contains the noise frequency identifiers for each detection region. The proportional relationship of the frequency stability coefficient ensures the consistency of noise assessment; high-frequency components typically have lower stability. The implementation of the frequency extraction module includes parallel processing of multiple frequency components, accelerated by a multi-core processor. The labeling of significant noise frequencies provides input to the error derivation module for error probability calculation.

[0048] The frequency stability coefficient is derived using a linear combination formula. The values ​​of the weighting factors α and β are based on a large amount of experimental data statistics. See Table 1 for typical configurations of the weighting factors in different frequency ranges.

[0049] Table 1: Frequency Stability Coefficient Weighting Factor Configuration Table

[0050] The weight factor configuration table is used to select the corresponding α and β values ​​based on the range of the frequency component being analyzed when calculating the frequency stability coefficient. The noise contamination probability calculation in the frequency extraction module is performed independently for each local window, with the local window division method consistent with the aforementioned embodiment. The preset threshold is set considering image quality requirements; a higher threshold is used for high-quality detection scenarios, and a medium threshold is used for regular scenarios. The labeling results of significant noise frequencies are stored in bitmap format, with each bit representing whether a frequency component is labeled. The proportional relationship of the frequency stability coefficient is achieved through a linear function, and the coefficient value is dynamically adjusted as the frequency changes. The processing flow of the frequency extraction module includes data preprocessing steps, smoothing and filtering the noise contamination probability to reduce the impact of random fluctuations. The weight factor configuration table for the frequency stability coefficient supports dynamic updates, and the factor values ​​can be adjusted based on the latest detection data during system operation. The significant noise frequency extraction algorithm includes connected component analysis, merging adjacent frequency components into noise regions. The adaptive adjustment of the preset threshold is based on the global statistical characteristics of the image, and the threshold value is automatically adjusted as the image contrast changes. The output format of the frequency extraction module is designed as a JSON structure for easy data exchange with other system modules. The calculation of the frequency stability coefficient uses vectorized instructions, processing multiple frequency components simultaneously to improve efficiency.

[0051] In the derivation of the frequency stability coefficient, the energy ratio calculation uses logarithmic transformation to enhance numerical stability, and the phase angle difference calculation uses angle normalization. The local window processing order of the frequency extraction module follows the spatial proximity principle, prioritizing adjacent windows to maintain continuity. The labeling of significant noise frequencies includes confidence assessment, with each labeled frequency accompanied by a reliability score. The method for determining the preset threshold includes sliding window statistics, adjusting the threshold based on local frequency characteristics. Performance optimization of the frequency extraction module includes memory access mode optimization to reduce cache misses. The proportionality relationship of the frequency stability coefficient is verified through unit testing, ensuring that the coefficient values ​​are within a reasonable range. The anomaly handling mechanism of the frequency extraction module includes filtering invalid frequency components, removing components exceeding the Nyquist frequency. The weight factor configuration table of the frequency stability coefficient includes boundary condition processing, with smooth transitions at frequency range boundaries to avoid abrupt changes. The storage of significant noise frequencies uses a sparse matrix format to save storage space. The application of preset thresholds includes a multi-threshold strategy, using different threshold values ​​for different frequency bands. The real-time performance of the frequency extraction module is guaranteed through pipeline design, overlapping data reading and calculation operations. Monitoring of the frequency stability coefficient includes numerical range checks, triggering recalculation upon detection of outliers.

[0052] The calibration process of the frequency extraction module uses standard test images containing known noise patterns. The optimization of the frequency stability coefficient's weighting factors employs a grid search method to find the optimal parameter combination. A visualization tool for significant noise frequencies supports frequency distribution display for debugging and analysis. The automatic adjustment algorithm for preset thresholds uses machine learning techniques to optimize threshold values ​​based on historical detection results. The frequency extraction module's expansion interface supports plug-in algorithms, allowing for the replacement of noise marker logic. The calculation of the frequency stability coefficient supports multi-precision modes, allowing selection of single-precision or double-precision operations as needed. The hardware acceleration scheme for the frequency extraction module uses FPGA to implement frequency domain filtering, improving processing speed. The weighting factor configuration table for the frequency stability coefficient is encrypted and stored to prevent unauthorized modification. The transmission protocol for significant noise frequencies uses a compressed format to reduce network bandwidth consumption. The user interface for preset thresholds provides interactive adjustment, allowing operators to manually fine-tune the threshold values. The frequency extraction module's log records detailed operation information, supporting fault diagnosis and performance analysis. Intermediate results in the calculation of the frequency stability coefficient are cached to avoid recalculating the same frequency components. The testing coverage includes boundary value testing to verify behavior under extreme frequency conditions; version management of the frequency stability coefficient weight factor configuration table, recording the history of each modification; export function for significant noise frequencies supporting multiple formats and compatible with third-party analysis tools; persistent storage of preset thresholds using configuration files, retaining settings after system restarts; power consumption optimization of the frequency extraction module through dynamic voltage and frequency adjustments to reduce energy consumption; and error propagation analysis to control the calculation accuracy of the frequency stability coefficient, ensuring the reliability of the final results. Multi-language support includes an internationalized interface to adapt to the needs of users in different regions; a backup mechanism for the frequency stability coefficient weight factor configuration table to prevent data loss; real-time monitoring of significant noise frequencies to display the current processing status, facilitating operator progress tracking; batch processing of preset thresholds supporting simultaneous adjustment of settings for multiple detection areas; fault-tolerant design of the frequency extraction module including redundant calculations, with primary and backup modules running simultaneously for result comparison; and interrupted recovery support for frequency stability coefficient calculation, allowing the processing to be paused and resumed. The maintenance tool provides automatic diagnostics to detect the health status of modules; the frequency stability coefficient weight factor configuration table verification program checks data integrity; the historical data comparison function of significant noise frequencies supports performance trend analysis; the intelligent recommendation of preset thresholds automatically suggests the optimal value based on image features; the upgrade mechanism of the frequency extraction module supports hot updates, allowing the algorithm to be replaced without stopping the system; the frequency stability coefficient calculation module is designed with isolation, so that a fault in a single frequency component does not affect the whole.

[0053] Example 5: See Figure 3The error derivation module calculates the defect detection error probability. Its input comes from the significant noise frequency data output by the frequency extraction module. The module collects all significant noise frequencies from the surface image data of each detection area; these frequencies are those marked by the frequency extraction module as potentially noise-containing frequency components. It then calculates a common subset of significant noise frequencies from all detection areas, achieved through set intersection operations, retaining only frequencies appearing in all detection areas. This common subset represents a system-level noise source. The frequencies in the common subset are used as key frequencies for subsequent error calculations. The module calculates the dispersion of the amplitude values ​​of the surface image data of each detection area and its adjacent areas at the key frequencies, using standard deviation to measure amplitude fluctuations. The average of these dispersion values ​​is taken as the noise-induced error for that detection area, smoothing out local fluctuations. The noise-induced errors for all detection areas are normalized using minimum-maximum scaling to map the error to the 0-1 range. The normalized result is used as the defect detection error probability; a higher probability value indicates lower defect detection reliability. The preset threshold is an adaptive threshold. The method for determining the preset threshold includes calculating the average and standard deviation of the probability of noise contamination for a specific frequency component within all local windows of a detection area; adding N times the standard deviation to the average value as the preset threshold for that frequency component, where N is a configurable constant greater than zero; the adaptive threshold adjustment improves the adaptability of noise detection based on the data distribution; the common subset search of the error derivation module reduces false alarms in individual detection areas, and key frequencies focus on common noise; the dispersion calculation uses a fast statistical algorithm to process a large amount of frequency data; normalization processing ensures the comparability of error probabilities, and the defect detection error probability serves as the final output to guide defect judgment.

[0054] The error derivation module prioritizes computational efficiency. Common subset selection uses a hash table to accelerate set operations, while determining key frequencies requires traversing the frequency markers of all detection regions. Discreteness calculation employs an incremental algorithm to avoid storing all amplitude values, saving memory. Normalization uses streaming computation to update minimum and maximum values ​​in real time. The output format of the defect detection error probability includes timestamps and region identifiers, supporting historical data tracing. The constant N for the adaptive threshold is determined experimentally, typically between 2 and 3, covering over 95% of the normal fluctuation range. The error derivation module's processing flow includes data validation steps, checking the validity of input frequency data and removing outliers. Common subset selection considers frequency tolerance, treating similar frequencies as the same frequency component. Key frequencies are stored using an ordered structure for rapid retrieval. Discreteness calculation supports multiple metrics, including variance and interquartile range. Normalization includes smoothing filtering to reduce the impact of abrupt changes. The release of defect detection error probabilities uses a message queue mechanism to ensure reliable data transmission. The error derivation module's parameter configuration includes the discreteness calculation method and normalization range, which affect the sensitivity of the error probability. The threshold for common subset selection is adjustable, allowing for the formation of common subsets even with missing detection areas. A limit on the number of key frequencies prevents excessive frequencies from causing computational complexity. The window size for discreteness calculation is configurable, balancing real-time performance and accuracy. The reference benchmark for normalization processing can be selected as global or local data. An alarm threshold for defect detection error probability can be set, triggering an alert when the threshold is exceeded. Performance optimization includes parallel computation of errors in each detection area using multi-core processors; pre-screening of common subsets using a Bloom filter to reduce computational load; a caching mechanism for key frequencies to avoid redundant calculations; vectorization of discreteness calculation to improve processing speed; a lookup table for normalization processing to accelerate the conversion process; and batch processing of defect detection error probabilities to support efficient computation. Anomaly handling includes strategies for compensating for missing data (interpolating neighboring regions when data in a detection area is invalid); handling empty sets in common subset extraction (using the union of all frequencies when no common frequencies exist); prioritizing key frequencies based on their frequency of occurrence; removing outliers from dispersion calculations using robust statistical methods; handling boundary cases in normalization (returning a uniform value when all values ​​are equal); and checking the reasonableness of defect detection error probabilities to ensure values ​​are within the 0-1 range. Calibration uses a standard test dataset containing samples with known error probabilities; accuracy verification of common subset extraction is achieved through manual annotation and comparison; analysis of the physical meaning of key frequencies combined with frequency characteristic explanations; accuracy testing of dispersion calculations uses synthetic data; consistency checks of normalization are performed across different datasets; and reliability assessment of defect detection error probabilities is conducted through comparison with actual defects.

[0055] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A surface defect detection system for coiled steel based on image denoising, characterized in that, The system includes: The image acquisition module is used to acquire surface image data of the inspection area of ​​the coiled steel. The fluctuation evaluation module is used to determine the texture fluctuation index of the detection area in each local window based on the gray-scale statistical characteristics of the surface image data of the detection area in the local window and the gray-scale statistical characteristics of the surface image data of adjacent detection areas in the corresponding local windows. The noise probability calculation module is used to combine the energy distribution differences, spectral phase shifts, and texture fluctuation indices of the surface image data of all local windows in adjacent detection areas in the frequency domain to evaluate the possibility of noise contamination in the surface image data of the detection area at each frequency component. The frequency extraction module is used to extract significant noise frequencies based on the noise contamination probability of all frequency components of the surface image data of the detection area. The error derivation module is used to calculate the probability of defect detection error based on the clustering characteristics of significant noise frequencies in all adjacent detection areas.

2. The image denoising-based surface defect detection system for coiled steel as described in claim 1, characterized in that, The fluctuation assessment module determines the texture fluctuation index by: dividing the surface image data into multiple local windows; for each local window, calculating the overlap area of ​​the gray-level histogram of the detection region and its adjacent detection regions within the window; and using the normalized value of the overlap area as the texture fluctuation index of the detection region in that local window.

3. The image denoising-based surface defect detection system for coiled steel as described in claim 2, characterized in that, The method for dividing the local window is as follows: the surface image data is decomposed into pyramid-style multi-scale decomposition to obtain image layers of different resolutions, and each image layer serves as a local window.

4. The image denoising-based surface defect detection system for coiled steel as described in claim 1, characterized in that, The noise probability calculation module assesses the likelihood of noise contamination by: for any two adjacent detection areas, using each local window of the surface image data of the preceding detection area as a reference window; identifying the set of pixels in the surface image data of the following detection area that are spatially aligned with the reference window to form a comparison window; calculating the energy ratio and phase angle difference between the reference window and the comparison window in the same frequency band, and deriving the frequency stability coefficient based on the energy ratio and phase angle difference; standardizing the quotient of the texture fluctuation index and the frequency stability coefficient, subtracting it from a fixed constant, and averaging it over all local windows to obtain the noise contamination likelihood of each frequency component.

5. The image denoising-based surface defect detection system for coiled steel as described in claim 4, characterized in that, The frequency stability coefficient is derived by multiplying the energy ratio and the phase angle difference by weighting factors and then adding them together. The sum is used as the frequency stability coefficient, which is directly proportional to both the energy ratio and the phase angle difference.

6. The image denoising-based surface defect detection system for coiled steel as described in claim 1, characterized in that, The frequency extraction module extracts significant noise frequencies by: calculating the noise contamination probability of each frequency component within all local windows of the surface image data of each detection area, and marking frequency components with a noise contamination probability exceeding a preset threshold as significant noise frequencies.

7. The image denoising-based surface defect detection system for coiled steel as described in claim 1, characterized in that, The error derivation module calculates the defect detection error probability by: collecting all significant noise frequencies of the surface image data of each detection area; obtaining a common subset of the significant noise frequencies of all detection areas, and using the frequencies in the common subset as key frequencies; calculating the dispersion of the amplitude values ​​of the surface image data of each detection area and its adjacent detection areas at the key frequencies, and taking the average of the dispersion as the noise-induced error of that detection area; normalizing the noise-induced errors of all detection areas, and using the normalized result as the defect detection error probability.

8. The image denoising-based surface defect detection system for coiled steel as described in claim 7, characterized in that, The preset threshold is an adaptive threshold, and its determination method includes: For a specific frequency component within all local windows of a detection area, calculate the average and standard deviation of all noise contamination probabilities. The average value plus N times the standard deviation is used as the preset threshold for the frequency component, where N is a configurable constant greater than zero.

9. The image denoising-based surface defect detection system for coiled steel as described in claim 3, characterized in that, The pyramid-style multi-scale decomposition employs the Gaussian pyramid algorithm, which specifically includes: continuously performing Gaussian smoothing and downsampling operations on the original surface image data to generate image representations at multiple scales, with each scale image representation serving as an analysis level.

10. The image denoising-based surface defect detection system for coiled steel as described in claim 2, characterized in that, The calculation of the overlap area of ​​the gray-level histogram of the detection area and its adjacent detection areas within the window includes: quantizing the gray-level values ​​within the local window into 256 levels; calculating the gray-level histograms of the detection area and the adjacent detection areas respectively, and normalizing the histograms so that the sum is 1; for each gray-level, taking the minimum value of the two histogram values; and summing the minimum values ​​of all gray-levels to obtain the overlap area.

Images