A method for synthesizing and restoring motion-blurred images of cold-rolled copper strips

Abstract

The present application belongs to the field of metallurgical rolling machine vision image processing, and discloses a cold-rolled copper strip motion blur image synthesis and restoration method. First, collect the on-site clear copper strip image, combine the rolling and the complete set of physical parameters of the camera to construct a blur scale model, superimpose the direction, scale and vibration disturbance to generate a simulation blur image, and construct a clear-fuzzy paired dataset; build a lightweight learnable wavelet multiscale deblurring network with high-frequency subband spatial attention, and strengthen the repair of high-frequency details such as copper strip edges and defects; use the paired dataset to train the restoration model DB_M and embed it into the production line vision system. The present application solves the problem of on-site paired sample scarcity, repairs the texture profile by frequency division, and balances the restoration accuracy and online reasoning speed with the lightweight structure, effectively improves the reliability of the strip width, run-out, defect and shape detection, and is suitable for various cold-rolled copper strip high-speed production lines and supports the whole-process intelligent vision detection as a pretreatment.

Description

Technical Field

[0001] This invention belongs to the field of intelligent monitoring and machine vision image processing technology in metallurgical rolling processes, and relates to an image synthesis and restoration method, specifically a method for synthesizing and restoring motion-blurred images of cold-rolled copper strip. Background Technology

[0002] Cold-rolled copper strip, with its excellent electrical conductivity, thermal conductivity, corrosion resistance, and plastic forming properties, has become a core raw material for semiconductor chip carriers, current collectors for new energy power batteries, 5G communication components, precision conductive components for aerospace, and heat dissipation pads for high-end equipment, with its market application scale continuing to expand. As downstream end products iterate towards ultra-thin specifications, micron-level dimensional tolerances, zero surface defects, and long-term reliable operation, unprecedentedly stringent standards have been imposed on online quality control throughout the entire cold-rolling production process. Microscopic defects such as scratches, pits, and peeling on the copper strip surface, lateral deviation of the strip, real-time bandwidth dimensions, and strip shape defects all require continuous online monitoring using high-precision vision systems. Manual sampling inspection, due to its low efficiency, high missed detection rate, and inability to achieve continuous monitoring across the entire strip, has been gradually replaced by automated vision inspection solutions. The machine vision-based online inspection system for cold-rolled copper strip has outstanding advantages such as non-contact measurement, millisecond-level imaging, simultaneous acquisition of massive texture information, and deep coupling with intelligent recognition algorithms. At present, it has been deployed on a large scale in various reversible cold rolling mills, undertaking multiple core quality control tasks such as surface defect detection, real-time strip width calculation, dynamic monitoring of deviation, strip shape and waveform recognition, and visual monitoring of mill operation status. It is a core component of modern intelligent manufacturing production lines for copper materials.

[0003] However, under the high-speed continuous conveyor belt operation of cold rolling mills, motion blur has become the core bottleneck restricting the accuracy and stability of vision inspection systems. The rolling speed of cold-rolled copper strip can generally reach 100~400m / min. Industrial area / line array cameras are limited by the power of the light source and the light-sensing capability of the sensor, and cannot shorten the exposure time indefinitely. Within the limited exposure window, the copper strip continuously moves along the rolling direction, and the fine textures, defect edges, and the boundaries on both sides of the strip will form obvious trailing shadows, directly causing image edge diffusion, loss of texture details, and a significant decrease in overall image contrast. This image degradation phenomenon can damage the output of downstream detection algorithms in multiple dimensions: In bandwidth and deviation measurement scenarios, blurred boundaries can cause pixel-level positioning offsets, which accumulate to produce tens of micrometer-level dimensional measurement errors, failing to meet the micrometer-level tolerance control requirements of high-end copper strips; in surface defect recognition tasks, the blurring of defect contours such as cracks and micro-indentations makes the algorithm prone to missed detections, and the pseudo-textures caused by blurring significantly increase the frequency of false alarms, increasing the workload of manual verification; for plate shape recognition modules such as edge waves, middle waves, and quarter waves, blurred images weaken the local waveform grayscale gradient changes, making it difficult for the model to accurately determine the wave amplitude and distribution range, causing the plate shape control system to lag, and thus leading to production accidents such as strip breakage and reduced yield. Therefore, motion blur image restoration is an essential link in establishing a high-precision visual inspection link in high-speed cold rolling production lines, and it is also a key engineering problem that urgently needs to be solved in the field of metallurgical vision.

[0004] Current image deblurring techniques can be broadly categorized into three main approaches: traditional non-blind deblurring, blind deblurring, and deep learning-based end-to-end restoration. Traditional non-blind deblurring methods require precise pre-observation of the global blur kernel, relying on deconvolution iterations to reconstruct the clear image. However, in industrial settings, rolling speeds and equipment vibrations are constantly changing, and the blur kernel lacks a fixed value, making this approach highly unsuitable. Blind deblurring algorithms can simultaneously estimate the blur kernel and the clear image, but the iterative computation is large and the convergence speed is slow, making it difficult to meet the real-time inference requirements of production lines. In recent years, deep learning technology, with its powerful nonlinear fitting capabilities, has made breakthroughs in deblurring tasks for general scenarios such as natural images and surveillance footage. By training networks with a large number of clear-blurred pairing samples, it can directly establish a mapping relationship from degraded images to clear images without explicitly solving for the blur kernel, possessing greater potential for scene adaptation. However, when directly transferring the general deep learning deblurring solution to the cold-rolled copper strip inspection scenario, two major pain points that cannot be overcome will be encountered. The existing Chinese invention patents with publication numbers CN112304958A and CN114627108A can fully demonstrate the inherent shortcomings of the current technology system.

[0005] Among them, CN112304958A presents an existing mature copper strip surface defect detection solution. This patent constructs a complete traditional visual workflow of "difference image defect initial judgment - adaptive Canny edge segmentation - defect feature extraction - feature matrix classification." It relies on gradient histograms to dynamically adjust the Canny threshold to improve the integrity of defect segmentation. The algorithm logic is designed only for copper strip images with clear imaging and no obvious motion blur, completely ignoring the motion blur degradation caused by high-speed rolling. The preprocessing stage of this patent only includes grayscale conversion and basic median filtering, which can only eliminate static salt-and-pepper noise and Gaussian noise, but cannot repair the edge and texture loss caused by motion blur. Once there is a blurred copper strip image in the input, the adaptive Canny operator cannot capture the incomplete defect boundary, resulting in distortion of geometric and texture feature extraction, and ultimately a significant drop in defect classification accuracy. At the same time, this patent does not involve a deep learning restoration framework at all, and there are no technical ideas for sample construction and network training for motion blur. It is only suitable for low-speed sampling offline detection and is completely unsuitable for high-speed continuous rolling online scenarios.

[0006] CN114627108A focuses on the standardized creation of a dataset for metal strip shape defects, providing a static dataset processing flow including on-site image acquisition, Gaussian filtering, Canny edge extraction, Hough line detection, and background mask removal. It can automatically remove invalid backgrounds such as roller conveyors and frames, and divide training and testing sample sets, providing standardized original image materials for the strip shape recognition model. However, this patent has a fundamental data deficiency: its dataset consists entirely of clear, unblurred copper strip images manually acquired at reduced speeds on the production line. It completely fails to incorporate imaging physical parameters such as rolling speed, camera exposure, object distance, and pixel size to construct a motion degradation model. Therefore, it cannot generate motion blur paired samples that correspond one-to-one with clear images, and cannot simulate the random blurring disturbances in scale and direction caused by mill vibration and strip deviation. Under a deep learning supervised training system, the lack of clear-blurred paired labeled data corresponding to real-world working conditions makes it impossible to train a dedicated restoration network for motion blur. This dataset can only serve the strip shape classification task under clear images and cannot solve the problem of image degradation on high-speed production lines. The two comparative documents cover two major sub-directions: copper strip defect detection and strip data set creation. However, both have their own technical limitations. Even if the two patented technical solutions are combined, it is still impossible to form a complete integrated technical chain of "fuzzy sample generation - deep learning restoration - defect / plate shape detection".

[0007] Besides the shortcomings of existing industrial vision patents, general-purpose open-source deep learning deblurring networks also struggle to adapt to the specific imaging characteristics of cold-rolled copper strip. Effective recognition information for cold-rolled copper strip images is highly concentrated on strip edges, minor defect boundaries, and surface micro-texture areas—all high-frequency image features. However, mainstream general-purpose deblurring networks rely solely on global spatial convolution for feature extraction, exhibiting weak ability to distinguish high-frequency details in horizontal, vertical, and diagonal directions. Even after restoration, minor defects and edges of the copper strip remain blurred. Furthermore, most general-purpose deblurring networks have a large number of parameters and high inference latency. Industrial edge control computers and embedded vision devices have limited computing power, making them unable to meet the real-time image throughput requirements of production lines and hindering their implementation in online preprocessing scenarios. In summary, existing technologies suffer from a triple deficiency: missing datasets, poor network structure adaptability to various scenarios, and insufficient inference efficiency, failing to meet the practical engineering needs of high-speed cold-rolled copper strip visual inspection.

[0008] In summary, the field of metallurgical machine vision urgently needs a complete technical solution for motion blur image synthesis and restoration specifically designed for cold-rolled copper strip production conditions. On the one hand, it is necessary to establish a quantitative motion fuzzing degradation model based on the complete set of physical parameters of rolling and imaging, superimpose random disturbances caused by mill vibration and strip deviation, and generate one-to-one corresponding clear-fuzzy paired samples in batches. This completely solves the industry pain point that it is extremely difficult to collect paired samples in the industrial field and fills the technical gap of no fuzzy samples in the CN114627108A static dataset. On the other hand, it is necessary to construct a dedicated wavelet domain defuzzing network that is adapted to high-frequency detail recovery of copper strip, lightweight and deployable online. This makes up for the shortcomings of CN112304958A, which lacks deep learning restoration and has weak anti-fuzzing ability of traditional algorithms. By frequency domain decomposition and high-frequency attention mechanism, the edge and defect texture restoration effect is enhanced in a targeted manner. At the same time, the model computation overhead is compressed. Finally, the trained restoration model is embedded into the front end of the existing automatic vision inspection system to improve the clarity of the input image from the source. This comprehensively improves the detection accuracy and system operation stability of the four core tasks of bandwidth measurement, deviation monitoring, surface defect recognition and strip shape detection, providing new technical support for online quality control of intelligent manufacturing of cold-rolled copper strip. Summary of the Invention

[0009] To address the shortcomings of existing technologies, this invention aims to provide a method for synthesizing and restoring motion-blurred images of cold-rolled copper strips. This method constructs a motion-blurred simulation model with perturbations based on imaging process parameters, generates multi-level paired datasets, designs a lightweight wavelet deblurring network with high-frequency attention, combines it with a dual-loss training model, and embeds it into the production line vision system to improve the accuracy and real-time performance of online detection for various types of copper strips.

[0010] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0011] A method for synthesizing and restoring motion-blurred images of cold-rolled copper strips, specifically including the following steps:

[0012] Step 1: Collect raw images of the cold-rolled copper strip production site and construct a raw image dataset of cold-rolled copper strip;

[0013] Step 2: Based on the imaging geometry during the high-speed movement of cold-rolled copper strip, construct a motion blur image synthesis model Mblur_M, and use the motion blur image synthesis model Mblur_M to generate motion blur images corresponding to sharp images, thus constructing a sharp-blur pairing dataset;

[0014] Step 3: Construct an intelligent de-motion blurring network for cold-rolled copper strip. The intelligent de-motion blurring network includes a feature embedding module, an encoding module, a learnable wavelet fusion module, a wavelet head decoding module, and a multi-scale image output module.

[0015] Step 4: Train, validate, and test the intelligent motion fuzzing network based on the clear-fuzzy pairing dataset to obtain the intelligent motion fuzzing model DB_M for cold-rolled copper strip;

[0016] Step 5: Embed the intelligent de-motion blur model DB_M of the cold-rolled copper strip into the automatic visual inspection system for cold rolling production, and restore the motion blur image of the collected cold-rolled copper strip in real time or near real time.

[0017] As a limitation of this invention, step 1 includes the following steps:

[0018] Step 11: Set up an industrial camera at the cold-rolled copper strip production site. The industrial camera is installed on the exit side of the rolling mill or obliquely above the plate roll to collect production images including the surface of the copper strip, the edge contour, or the plate roll area.

[0019] Step 12: Use a computer to crop the acquired video stream or continuous images, set the image pixel size to H×W, and crop a total of n images as the raw image dataset Raw_dataset;

[0020] Step 13: Select images from Raw_dataset that are clear, have intact copper band regions, and have no obvious exposure anomalies to construct the clear source image dataset RawClear_dataset.

[0021] As another limitation of the present invention, the motion blur image synthesis model Mblur_M in step 2 determines the motion blur scale in the following way: Let the rolling speed be... The camera exposure time is The lens focal length is The distance from the target to the camera is The pixel size is The length of the motion blur kernel is Under the condition that the direction of copper strip motion is approximately parallel to the camera image plane, the length of the motion blur kernel... satisfy: ;in, Used to characterize the pixel displacement on the image plane caused by the movement of the copper strip during camera exposure time; when Using m / min, Using ms, and Using mm, When using μm, the length of the motion blur kernel is expressed as: .

[0022] As a further limitation of the present invention, the motion blur image synthesis model Mblur_M includes the following steps:

[0023] Step 21: Select any clear image from the RawClear_dataset. And read or set the rolling speed corresponding to the image. Exposure time ,focal length Object distance and pixel size ;

[0024] Step 22: According to Computing the fundamental fuzzy scale ;when Using m / min, Using ms, and Using mm, When using μm, the basic fuzzy scale Represented as:

[0025] Step 23: Set the direction of movement and in Introducing random directional bias on the basis , obtain the fuzzy kernel direction ;

[0026] Step 24: At the basic fuzzy scale Introducing random scale perturbation The final fuzzy scale is obtained. ;

[0027] Step 25: According to and Generate motion blur kernel ;

[0028] Step 26: Clear the image With motion fuzzy kernel Convolution is performed, and random noise used to simulate mill vibration or transverse perturbations in the strip is superimposed. To obtain a synthetic motion-blurred image ,in, This represents the convolution operation. This indicates random noise or spatial disturbance noise.

[0029] As a further limitation of the present invention, random direction deviation For a random value within a preset angle range, in a preferred embodiment, the following is satisfied: The random scale perturbation For a random value within a preset pixel range, in a preferred embodiment, the following is satisfied: The random noise Transverse random noise or spatial disturbance noise is used to simulate non-ideal imaging factors caused by mill vibration, strip deviation, and micro-vibration of the acquisition system in the cold rolling field.

[0030] As a third limitation of the present invention, step 2 further includes adjusting the fuzzy scale. or rolling speed The synthesized motion-blurred image is divided into multiple blur levels, which include at least slight blur. Moderate blur and severe blur Based on fuzzy scale The fuzzy level classification satisfies: , , ;in, and These are the thresholds between mild and moderate blur, and between moderate and severe blur, respectively; in a preferred embodiment, the threshold is: , Correspondingly, based on rolling speed The fuzzy level classification is as follows: , , .

[0031] As a fourth limitation of the present invention, the intelligent motion-de-blurring network in step 3 includes:

[0032] The feature embedding module is used to map the input motion-blurred image into shallow features;

[0033] The encoding module, consisting of multiple simple encoding blocks (SEBs), is used for progressive downsampling and extraction of multi-scale features; the learnable wavelet fusion module (WFB) is used to transform the encoded features of adjacent scales to the wavelet domain and perform cross-scale fusion; the wavelet head decoding module (WHB) is used to progressively upsample the fused features and reconstruct the image; and the multi-scale output module is used to output the restored images at different scales and participate in multi-scale supervision during the training phase.

[0034] As a further limitation of the present invention, both the learnable wavelet fusion module WFB and the wavelet head decoding module WHB include learnable wavelet nodes LWN, wherein the learnable wavelet nodes LWN include forward wavelet convolution, deep feature transformation, and backward wavelet convolution; the forward wavelet convolution is composed of a one-dimensional low-frequency filter. and one-dimensional high-frequency filters Constructing a two-dimensional wavelet convolution kernel using outer product: , , , ,in, For low-frequency subband convolution kernels, For horizontal high-frequency subband convolution kernels, It is a vertical high-frequency subband convolution kernel. The four two-dimensional wavelet convolution kernels are concatenated to form a diagonal high-frequency subband convolution kernel. The complete wavelet convolution kernel is obtained by concatenating these four kernels. For the input feature map It is mapped to the wavelet domain through grouped convolution: And four types of sub-band features were obtained: The one-dimensional low-frequency filter and one-dimensional high-frequency filters These are learnable parameters that can be adaptively updated based on the image features of cold-rolled copper strip during network training.

[0035] As a further limitation of the present invention, a spatial attention mechanism guided by high-frequency subbands can be set in the learnable wavelet node (LWN), wherein the spatial attention mechanism is based on the horizontal high-frequency subbands. and vertical high-frequency subband Generate spatial attention weights. First, concatenate the horizontal and vertical high-frequency subbands: Then, spatial attention weights are generated using convolutional mapping and the sigmoid function: ,in, This represents the Sigmoid activation function. Indicates the kernel size as The convolution operation; the spatial attention weights After size alignment, the spatial domain features applied to the wavelet domain features are used to obtain enhanced features: ,in, This indicates element-wise multiplication; through the above method, the network focuses on the edge contour, surface texture, and defect boundary areas of the cold-rolled copper strip.

[0036] As a further limitation of the present invention, the depth feature transformation in the learnable wavelet node (LWN) adopts a lightweight structure, which replaces the 7×7 depth-separable convolution with a 3×3 depth-separable convolution and sets a 1×1 convolution, a GELU activation function, a PixelShuffle upsampling branch or an element-wise multiplication fusion branch to reduce the number of model parameters and inference time.

[0037] As a fifth limitation of the present invention, step 4 includes the following steps:

[0038] Step 41: Set the model training parameters, which include input image size, batch size, optimizer, initial learning rate, learning rate decay strategy, maximum number of iterations, and weight saving interval;

[0039] Step 42: Input the motion-blurred images in the training set into the intelligent de-motion blurring network, and output multi-scale restored images;

[0040] Step 43: According to the first Restored images at various scales With corresponding clear label images Calculate the multi-scale restoration loss: ,in, Indicates the number of output scales. Indicates the first Loss weights at each scale, The pixel-level restoration loss function is defined as follows: ;

[0041] Step 44: Calculate the wavelet constraint loss based on the perfect reconstruction constraints of the wavelet filter. Let the forward wavelet filter be... , The inverse wavelet filter is , Then the perfect reconstruction constraint is satisfied: The corresponding wavelet constraint loss is: ;

[0042] Step 45: Update the network parameters according to the total loss function, which is: ,in, The weights are set for wavelet-constrained loss.

[0043] Step 46: Select the model weights with the best overall performance on the validation set as the intelligent motion fuzzing model DB_M for cold-rolled copper strip.

[0044] As a sixth limitation of the present invention, step 5 includes: inputting the cold-rolled copper strip image acquired in real time by an industrial camera into the intelligent motion blur removal model DB_M of the cold-rolled copper strip, outputting the restored clear image, and using the clear image as input for strip width measurement, deviation detection, semantic segmentation, surface defect detection, or plate shape visual inspection models; specifically, let the motion blur image acquired in real time by the industrial camera be... The mapping function of the intelligent motion-de-fuzzy model DB_M for cold-rolled copper strip is: The restored clear image Represented as: The restored clear image Input downstream automatic visual inspection model The test results were obtained. The detection results This includes at least one of the following: strip edge position, strip width, deviation amount, semantic segmentation result, surface defect detection result, or plate shape visual inspection result.

[0045] By adopting the above-described technical solution, the beneficial effects achieved by this invention compared to the prior art are as follows:

[0046] (1) The motion blur image synthesis model Mblur_M proposed in this invention introduces on-site physical parameters such as rolling speed, exposure time, focal length, object distance, and pixel size into the image degradation modeling process, and can establish a quantitative mapping relationship from cold rolling process parameters to blur scale. This model can generate motion blur images that are strictly paired with sharp source images, effectively solving the problem of difficulty in obtaining sharp-blur paired samples in industrial settings.

[0047] (2) In the process of generating blurred images, the present invention further introduces random perturbation of motion direction, random perturbation of fuzzy scale and vibration noise to simulate non-ideal factors such as strip deviation, mill vibration and micro-vibration of acquisition system, so that the synthesized image is closer to the motion fuzzing degradation characteristics of the real cold rolling production site.

[0048] (3) The intelligent de-motion blurring network constructed in this invention can explicitly decompose the low-frequency structural information and high-frequency edge texture information in the image through learnable discrete wavelet transform, so that the network can differentiate the degradation features of different frequencies and directions, which is beneficial to recover the edge contour, surface texture and defect boundary of cold-rolled copper strip.

[0049] (4) The present invention constructs a spatial attention mechanism on the basis of the high frequency subband of the wavelet domain, so that the model pays more attention to the strip boundary and texture details during the restoration process, improves the edge sharpness and structural integrity of the restored image, thereby improving the reliability of subsequent strip position detection, surface defect detection and plate shape visual detection.

[0050] (5) This invention reduces the number of network parameters and inference time by using lightweight wavelet nodes, thereby improving the model’s online application potential while ensuring restoration quality. It can be used for real-time or near-real-time image preprocessing in automatic visual inspection systems for cold-rolled copper strip production lines.

[0051] In summary, this invention establishes a quantitative fuzzy scale model based on rolling and imaging physical parameters, superimposes multiple random perturbations to replicate the degradation effect of real-world images, and constructs a complete paired dataset. At the same time, it designs a lightweight learnable wavelet restoration network that integrates high-frequency attention to specifically repair copper strip edges and defect textures by frequency division, balancing restoration accuracy and real-time inference efficiency. When embedded in the detection system, it significantly improves the reliability of various online visual inspection tasks.

[0052] This invention is applicable to various cold-rolled copper strip continuous rolling production lines, and is used for real-time deblurring and preprocessing of images under high-speed motion conditions, supporting intelligent visual inspection of bandwidth, deviation, surface defects, and strip shape throughout the entire process. Attached Figure Description

[0053] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0054] Figure 1 This is a schematic diagram of the overall process of the method for synthesizing and restoring motion blur images of cold-rolled copper strip according to the present invention;

[0055] Figure 2 This is a geometrical schematic diagram illustrating the image plane displacement caused by the movement of the strip during the camera exposure time of the present invention.

[0056] Figure 3 This is a schematic diagram of the overall structure of the intelligent motion fuzzy removal network for cold-rolled copper strip according to the present invention;

[0057] Figure 4 This is a schematic diagram illustrating the construction of the learnable wavelet convolution kernel and the group convolution processing of the present invention;

[0058] Figure 5 This is a schematic diagram of the high-frequency subband guided spatial attention and lightweight LWN module structure of the present invention;

[0059] Figure 6 This is a schematic diagram of the WFB and WHB module structures of the present invention;

[0060] Figure 7 This is a schematic diagram of the SEB module structure of the present invention. Detailed Implementation

[0061] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustrative and understanding purposes only and are not intended to limit the scope of the invention.

[0062] Example: A method for synthesizing and restoring motion-blurred images of cold-rolled copper strip

[0063] like Figures 1 to 7 As shown, this embodiment relates to a method for synthesizing and restoring motion-blurred images of cold-rolled copper strips, specifically including the following steps:

[0064] Step 1: Collect raw images of the cold-rolled copper strip production site and construct a raw image dataset of cold-rolled copper strip;

[0065] Step 11: Set up an industrial camera at the cold-rolled copper strip production site. The industrial camera is installed on the exit side of the rolling mill or obliquely above the plate roll to collect production images including the surface of the copper strip, the edge contour, or the plate roll area.

[0066] Step 12: Use a computer to crop the acquired video stream or continuous images, set the image pixel size to H×W, and crop a total of n images as the raw image dataset Raw_dataset;

[0067] Step 13: Select images from Raw_dataset that are clear, have intact copper band regions, and have no obvious exposure anomalies to construct the clear source image dataset RawClear_dataset.

[0068] Step 2: Based on the imaging geometry during the high-speed movement of cold-rolled copper strip, construct a motion blur image synthesis model Mblur_M, and use the motion blur image synthesis model Mblur_M to generate motion blur images corresponding to sharp images, thus constructing a sharp-blur pairing dataset;

[0069] The motion blur image synthesis model Mblur_M determines the motion blur scale in the following way: Let the rolling speed be... The camera exposure time is The lens focal length is The distance from the target to the camera is The pixel size is The length of the motion blur kernel is Under the condition that the direction of copper strip motion is approximately parallel to the camera image plane, the length of the motion blur kernel... satisfy: ;in, Used to characterize the pixel displacement on the image plane caused by the movement of the copper strip during camera exposure time; when Using m / min, Using ms, and Using mm, When using μm, the length of the motion blur kernel is expressed as: ;

[0070] The motion-blurred image synthesis model Mblur_M includes the following steps:

[0071] Step 21: Select any clear image from the RawClear_dataset. And read or set the rolling speed corresponding to the image. Exposure time ,focal length Object distance and pixel size ;

[0072] Step 22: According to Computing the fundamental fuzzy scale ;when Using m / min, Using ms, and Using mm, When using μm, the basic fuzzy scale Represented as:

[0073] Step 23: Set the direction of movement and in Introducing random directional bias on the basis , obtain the fuzzy kernel direction ;

[0074] Random direction deviation For a random value within a preset angle range, in a preferred embodiment, the following is satisfied: The random scale perturbation For a random value within a preset pixel range, in a preferred embodiment, the following is satisfied: The random noise Transverse random noise or spatial disturbance noise is used to simulate non-ideal imaging factors caused by mill vibration, strip deviation, and micro-vibration of the acquisition system in the cold rolling field.

[0075] Step 24: At the basic fuzzy scale Introducing random scale perturbation The final fuzzy scale is obtained. ;

[0076] According to fuzzy scale or rolling speed The synthesized motion-blurred image is divided into multiple blur levels, which include at least slight blur. Moderate blur and severe blur Based on fuzzy scale The fuzzy level classification satisfies: , , ;in, and These are the thresholds between mild and moderate blur, and between moderate and severe blur, respectively; in a preferred embodiment, the threshold is: , Correspondingly, based on rolling speed The fuzzy level classification is as follows: , , ;

[0077] Step 25: According to and Generate motion blur kernel ;

[0078] Step 26: Clear the image With motion fuzzy kernel Convolution is performed, and random noise used to simulate mill vibration or transverse perturbations in the strip is superimposed. To obtain a synthetic motion-blurred image ,in, This represents the convolution operation. This indicates random noise or spatial disturbance noise.

[0079] Step 3: Construct an intelligent de-motion blurring network for cold-rolled copper strip. The intelligent de-motion blurring network includes a feature embedding module, an encoding module, a learnable wavelet fusion module, a wavelet head decoding module, and a multi-scale image output module.

[0080] The intelligent motion blur removal network includes: a feature embedding module, which maps the input motion-blurred image into shallow features;

[0081] Motion blur degradation model: This embodiment constructs a motion blur image dataset for cold-rolled copper strip. Specific steps are as follows:

[0082] Step 31: Read clear images from RawClear_dataset ;

[0083] Step 32: Setting and The corresponding physical parameters, params, include at least the rolling speed. Exposure time Lens focal length Object distance and pixel size ;

[0084] Step 33: According to the formula Calculate the basic fuzzy scale;

[0085] Step 34: Set random direction deviation and random scale perturbation The final fuzzy kernel direction is obtained. and final fuzzy scale ;

[0086] Step 35: According to and Generate motion blur kernel ;

[0087] Step 36: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] and Perform convolution and add random noise To obtain a synthetic motion-blurred image ;

[0088] Step 37: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] As a clear label image, As corresponding motion-blurred images, they constitute a set of sharp-blurred paired samples;

[0089] Step 38: Repeat steps 31 to 37 for all clear images in the RawClear_dataset to obtain the clear-blur paired dataset SynBlur_dataset;

[0090] To ensure the synthesized data covers different rolling speed conditions, this embodiment sets the rolling speed range to 100 m / min to 400 m / min, and classifies the motion-blurred image into mild blur S1, moderate blur S2, and severe blur S3 based on the rolling speed or blur scale. In a preferred embodiment, the blur level classification is shown in Table 1:

[0091] Table 1:

[0092] In a preferred embodiment, to avoid data leakage, the training set, validation set, and test set are first divided at the level of the clear source image, and then motion-blurred images of corresponding blur levels are generated respectively, so that different blurred versions of the same clear source image will not appear in the training set and test set at the same time.

[0093] The encoding module, consisting of multiple simple encoding blocks (SEBs), is used for progressive downsampling and extraction of multi-scale features; the learnable wavelet fusion module (WFB) is used to transform the encoded features of adjacent scales to the wavelet domain and perform cross-scale fusion; the wavelet head decoding module (WHB) is used to progressively upsample the fused features and reconstruct the image; and the multi-scale output module is used to output the restored images at different scales and participate in multi-scale supervision during the training phase.

[0094] Both the learnable wavelet fusion module (WFB) and the wavelet head decoding module (WHB) include learnable wavelet nodes (LWNs), which include forward wavelet convolution, deep feature transformation, and backward wavelet convolution. The forward wavelet convolution is performed by a one-dimensional low-frequency filter. and one-dimensional high-frequency filters Constructing a two-dimensional wavelet convolution kernel using outer product: , , , ,in, For low-frequency subband convolution kernels, For horizontal high-frequency subband convolution kernels, It is a vertical high-frequency subband convolution kernel. The four two-dimensional wavelet convolution kernels are concatenated to form a diagonal high-frequency subband convolution kernel. The complete wavelet convolution kernel is obtained by concatenating these four kernels. For the input feature map It is mapped to the wavelet domain through grouped convolution: And four types of sub-band features were obtained: The one-dimensional low-frequency filter and one-dimensional high-frequency filters These are learnable parameters that can be adaptively updated based on the image features of cold-rolled copper strip during network training.

[0095] Learnable wavelet nodes (LWNs) can set up a spatial attention mechanism guided by high-frequency subbands, the spatial attention mechanism being based on the horizontal high-frequency subbands. and vertical high-frequency subband Generate spatial attention weights. First, concatenate the horizontal and vertical high-frequency subbands: Then, spatial attention weights are generated using convolutional mapping and the sigmoid function: ,in, This represents the Sigmoid activation function. Indicates the kernel size as The convolution operation; the spatial attention weights After size alignment, the spatial domain features applied to the wavelet domain features are used to obtain enhanced features: ,in, This indicates element-wise multiplication; through the above method, the network focuses on the edge contour, surface texture, and defect boundary areas of the cold-rolled copper strip.

[0096] The deep feature transformation in the learnable wavelet node (LWN) adopts a lightweight structure, which replaces the 7×7 depthwise separable convolution with a 3×3 depthwise separable convolution and sets 1×1 convolution, GELU activation function, PixelShuffle upsampling branch or element-wise multiplication fusion branch to reduce the number of model parameters and inference time.

[0097] Step 4: Train, validate, and test the intelligent motion fuzzing network based on the clear-fuzzy pairing dataset to obtain the intelligent motion fuzzing model DB_M for cold-rolled copper strip;

[0098] Step 41: Set the model training parameters, which include input image size, batch size, optimizer, initial learning rate, learning rate decay strategy, maximum number of iterations, and weight saving interval;

[0099] Step 42: Input the motion-blurred images in the training set into the intelligent de-motion blurring network, and output multi-scale restored images;

[0100] Step 43: According to the first Restored images at various scales With corresponding clear label images Calculate the multi-scale restoration loss: ,in, Indicates the number of output scales. Indicates the first Loss weights at each scale, The pixel-level restoration loss function is defined as follows: ;

[0101] Step 44: Calculate the wavelet constraint loss based on the perfect reconstruction constraints of the wavelet filter. Let the forward wavelet filter be... , The inverse wavelet filter is , Then the perfect reconstruction constraint is satisfied: The corresponding wavelet constraint loss is: ;

[0102] Step 45: Update the network parameters according to the total loss function, which is: ,in, The weights are set for wavelet-constrained loss.

[0103] Step 46: Select the model weights with the best overall performance on the validation set as the intelligent motion fuzzing model DB_M for cold-rolled copper strip.

[0104] Step 5: Embed the intelligent de-motion blur model DB_M of the cold-rolled copper strip into the automatic visual inspection system for cold rolling production, and restore the motion blur image of the collected cold-rolled copper strip in real time or near real time.

[0105] The cold-rolled copper strip image acquired in real time by an industrial camera is input into the intelligent motion blur removal model DB_M for cold-rolled copper strip, and the restored clear image is output. This clear image is then used as input for models such as strip width measurement, deviation detection, semantic segmentation, surface defect detection, or sheet shape visual inspection. Specifically, let the motion-blurred image acquired in real time by the industrial camera be... The mapping function of the intelligent motion-de-fuzzy model DB_M for cold-rolled copper strip is: The restored clear image Represented as: The restored clear image Input downstream automatic visual inspection model The test results were obtained. The detection results This includes at least one of the following: strip edge position, strip width, deviation amount, semantic segmentation result, surface defect detection result, or plate shape visual inspection result.

[0106] In this embodiment, an industrial area array camera is installed at the cold-rolled copper strip production site. The camera is mounted on the mill exit side or obliquely above the strip roll. The camera's field of view covers the surface area of ​​the copper strip, the edge area of ​​the copper strip, and the background structure area relevant to subsequent visual inspection tasks. The acquisition system includes an industrial camera, lens, light source, mounting bracket, image acquisition card or network transmission unit, and industrial computer.

[0107] During image acquisition, industrial cameras capture video or continuous images of cold-rolled copper strip production, and industrial computers extract frames. The number of frames to be extracted is set to... Image size is A total of The images are used as the raw image dataset, Raw_dataset.

[0108] To ensure the quality of subsequent motion-blurred image synthesis, clear images are selected from the Raw_dataset. The selection criteria include: complete copper strip main area, clear edge contours, visible surface texture, no severe overexposure or underexposure, no obvious fogging, and no large-area obstructions. The selected clear images constitute the RawClear_dataset.

[0109] In a preferred embodiment, the acquisition system is deployed on an 800mm four-roll reversible cold rolling mill, and the acquired copper strip width covers a range of 400mm to 680mm. The images include different alloy grades, different surface conditions, and different width specifications to improve the data set's coverage of actual production conditions.

[0110] Motion blur in the image acquisition of cold-rolled copper strip surfaces is mainly caused by the high-speed movement of the copper strip during the camera exposure time. Let the camera exposure time be... The copper strip rolling speed is Then during the exposure time Within the image, any point on the surface of the copper strip is displaced along the direction of motion. The projection of this point onto the imaging sensor is no longer a stationary image point, but forms a continuous trajectory within the exposure time, ultimately appearing as a trailing shadow along the rolling direction in a single frame image.

[0111] Let the distance between the target and the camera be... The lens focal length is The pixel size is The camera exposure time is The rolling speed is During the exposure time Inside, the object points on the copper strip surface undergo actual displacement along the direction of motion. The actual displacement can be expressed as: Based on the similarity imaging relationship, the actual displacement The corresponding displacement on the image plane for: Image plane displacement The corresponding pixel length is the motion blur kernel length. Then we have: Thus, the length of the motion fuzzy kernel is obtained. The relationship between rolling speed, exposure time, focal length, object distance, and pixel size is as follows: ;

[0112] In the above formula, The length of the motion blur kernel, in pixels; For rolling speed; For camera exposure time; The focal length of the lens; The distance from the target to the camera; For pixel size. When , , , and When using a unified unit, the formula can be used directly. Calculate the length of the motion fuzzy kernel.

[0113] In the cold-rolled copper strip production site, to facilitate calculations based on actual processes and camera parameters, when Using m / min, Using ms, and Using mm, When using μm, the length of the motion fuzzy kernel after unit conversion is considered. Represented as: In a preferred embodiment, when =25、 =800mm =3.45μm, At 5ms, the motion fuzzy kernel length and rolling speed satisfy the following relationship: ;

[0114] In actual cold rolling mills, the direction of strip movement is not entirely constant. Mill vibration, strip misalignment, tension fluctuations, and micro-vibrations during camera installation all contribute to a degree of randomness in the image degradation process. Therefore, this invention introduces a perturbation term based on the fundamental fuzzy scale and fundamental motion direction. Let the fundamental fuzzy scale be... ,but: At the basic fuzzy scale Introducing random scale perturbation The final fuzzy scale is obtained. : Meanwhile, let the basic direction of motion be... Based on this, a random directional bias is introduced. The final fuzzy kernel direction is obtained. : ;

[0115] in, The final motion blur kernel length, For scale perturbation terms, Based on the direction of motion, This refers to the directional perturbation term. Random values ​​within a preset angle range can be selected, and in a preferred embodiment, the following conditions are met: The scale perturbation term Random values ​​within a preset pixel range can be selected, and in a preferred embodiment, the following conditions are met: ;

[0116] Based on the final fuzzy scale and the final fuzzy kernel direction Generate motion blur kernel The motion fuzzy kernel For along direction , length is Linear convolution kernels. This will sharpen the image. With motion fuzzy kernel Convolution is performed, and random noise is superimposed to simulate mill vibration, strip transverse perturbation, and acquisition system noise. To obtain a synthetic motion-blurred image : ,in, This represents the convolution operation. This indicates random noise or spatial disturbance noise.

[0117] This embodiment constructs an intelligent motion blur removal network for cold-rolled copper strip. The network adopts a single-input, multi-output, multi-scale restoration architecture. The input is a motion-blurred image, and the output is a restored image at one or more scales. During the training phase, outputs at different scales participate in supervision; during the inference phase, the highest resolution output is selected as the final restoration result.

[0118] The network comprises a feature embedding module, an encoding module, a learnable wavelet fusion module (WFB), a wavelet head decoding module (WHB), and a multi-scale image output module. The feature embedding module converts the input image into shallow features. The encoding module consists of multiple simple coding blocks (SEBs), each SEB including LayerNorm, 1×1 convolution, depthwise convolution, GELU activation function, channel attention, and residual connections, used to extract spatial features at different scales.

[0119] The learnable wavelet fusion module (WFB) is used to fuse features from different coding scales. The WFB internally incorporates learnable wavelet nodes (LWNs), which transform spatial domain features to the wavelet domain, obtaining low-frequency structural information and high-frequency detail information from different directions. Feature enhancement and fusion are then performed in the wavelet domain, followed by conversion back to the spatial domain via inverse wavelet convolution.

[0120] The wavelet head decoding module (WHB) is used to progressively restore feature resolution and generate a restored image. WHB also includes a low-frequency wavelet head (LWN) to enhance the decoding stage's ability to detect frequency domain information and directional edges.

[0121] The learnable wavelet node (LWN) includes forward wavelet convolution, wavelet domain deep feature transform, and backward wavelet convolution. Forward wavelet convolution first defines a one-dimensional low-frequency filter. and one-dimensional high-frequency filters And four two-dimensional wavelet convolution kernels are generated through outer product: , , , ;

[0122] in, Corresponding low-frequency subband , Corresponding horizontal high-frequency sub-band , Corresponding vertical high-frequency sub-band , Corresponding to the diagonal high-frequency subband The four two-dimensional wavelet convolution kernels mentioned above are concatenated into a complete wavelet convolution kernel. : ;

[0123] wavelet convolution kernel The input features are applied using grouped convolution, mapping the input features to a subset containing... , , and Feature diagrams of four types of subbands. Subbands retain the overall brightness and structural information of the image. and The sub-bands respond to edge information from different directions. Subband response diagonal details and noise information.

[0124] Since the wavelet filter parameters can be learned during training, the network can adaptively adjust the frequency decomposition method according to the distribution of cold-rolled copper strip image data, thereby enhancing its ability to model the directional degradation features of motion blur.

[0125] Visual inspection of cold-rolled copper strip is highly sensitive to edge and texture details. The strip edge profile directly affects the accuracy of width detection and deviation measurement, while defect boundaries and surface texture directly affect the surface defect detection effect. Motion blur mainly weakens these high-frequency details; therefore, this embodiment introduces a high-frequency sub-strip guided spatial attention mechanism in LWN.

[0126] Specifically, the horizontal high-frequency subband LH and vertical high-frequency subband HL output by LWN contain a large amount of edge and texture orientation information. After concatenating or fusing LH and HL, they are input into the spatial attention generation unit to obtain the spatial attention weights. : , ;

[0127] Spatial attention weights After size alignment, the enhanced features are obtained by element-wise multiplication with the wavelet domain features or the spatial domain features before and after inverse wavelet convolution: In this way, the network can enhance its attention to edge contours, surface textures, defect boundaries, and local abrupt change regions.

[0128] Considering the high requirements for online processing speed in the automatic visual inspection system for cold-rolled copper strip, this embodiment makes a lightweight improvement to the LWN module. Specifically, the 7×7 depthwise separable convolution in the original LWN module is replaced with a 3×3 depthwise separable convolution to reduce the amount of convolution computation; at the same time, the 1×1 convolution is retained for channel dimension transformation, and the GELU activation function is retained to enhance nonlinear expression capabilities.

[0129] Furthermore, a PixelShuffle upsampling branch and a 1×1 convolution branch are set in LWN to improve feature resolution and enhance local detail recovery capabilities under low computational requirements. The lightweight LWN module fuses the outputs of the main branch and auxiliary branch by element-wise multiplication or addition, maintaining good restoration quality while reducing the number of parameters and inference time.

[0130] The intelligent motion blur removal network is trained using the sharp-blurred pairing dataset constructed above. During training, the input is a synthetic motion-blurred image, and the label is the corresponding sharp source image.

[0131] The network outputs restored images at multiple scales. Let the... The restored image at each scale is The corresponding clear label image is The scale weight is Then multi-scale recovery loss Represented as: ;

[0132] in, Indicates the number of output scales. Indicates the first Loss weights at each scale, The L1 norm is represented. In other implementations, the multi-scale restoration loss may also be the L2 loss, Charbonnier loss, structural similarity loss, or a combination thereof.

[0133] To prevent the learnable wavelet filter from degenerating into a regular convolutional kernel, this embodiment further introduces wavelet constraint loss. Wavelet constraint loss is used to constrain low-frequency and high-frequency filters to meet perfect reconstruction conditions, enabling forward and backward wavelet convolution to have stable frequency domain decomposition and reconstruction capabilities.

[0134] Total network loss Represented as:

[0135] in, The wavelet-constrained loss weights are used. During training, an optimizer iteratively updates the network parameters, saving the model weights after each preset number of iterations. Image quality metrics and model real-time performance metrics are evaluated on the validation set. Finally, the weight file with the best overall performance on the validation set is selected as the intelligent motion blur removal model DB_M for cold-rolled copper strip.

[0136] After model training is complete, the performance of DB_M can be evaluated using a test set. Evaluation metrics include subjective visual quality, peak signal-to-noise ratio (PSNR), structural similarity (SSIM), number of model parameters, average inference time per image, and performance on downstream visual detection tasks.

[0137] In a preferred embodiment, the image restored from DB_M is input into a semantic segmentation model for strip region segmentation and edge localization; alternatively, the restored image can be input into a surface defect detection model for detecting defects such as cracks, scratches, indentations, and oxide spots. Because the restored image possesses clearer edge, texture, and structural information, it enhances the ability of subsequent visual detection models to identify strip boundaries and defect regions.

[0138] During online deployment, industrial cameras acquire images of cold-rolled copper strip in real time and transmit the raw images to an industrial computer or edge computing device. DB_M performs motion blur removal and restoration on the input images, outputting a clear image. This clear image is then input into subsequent visual inspection modules, including but not limited to strip width measurement modules, deviation detection modules, semantic segmentation modules, surface defect detection modules, and strip shape visual inspection modules.

[0139] When the production line speed is high, a buffered online processing method can be adopted. This involves industrial cameras continuously acquiring images and writing them to a buffer queue. DB_M then restores the images in the queue order, and subsequent detection modules read images from the restored image queue for inspection. This method can improve system throughput and is suitable for continuous cold-rolled copper strip production scenarios.

[0140] This embodiment provides a preferred parameter setting, but the invention is not limited thereto. Camera pixel size It can be set to 3.45μm, lens focal length It can be set to 25mm, target distance. It can be set to 800mm, and the exposure time... It can be set to 5ms, basic motion direction. It can be set according to the rolling direction, with random directional deviation. It can be set from -10° to 10°.

[0141] When the rolling speed range is set to 100m / min to 400m / min, it can be based on The fuzzy kernel length is calculated, and the fuzziness levels are further classified as mild, moderate, and severe.

[0142] During model training, input images can be uniformly cropped or scaled to a preset size, such as 800×300 pixels; the optimizer can be AdamW; the initial learning rate can be set to 0.0002; the batch size can be set according to the video memory capacity; the maximum number of iterations can be set to 40000; and the model weights are saved once after each preset number of iterations.

[0143] It should be noted that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art can still modify the technical solutions described in the above embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for synthesizing and restoring motion-blurred images of cold-rolled copper strip, characterized in that: Specifically, the following steps are included: Step 1: Collect raw images of the cold-rolled copper strip production site and construct a raw image dataset of cold-rolled copper strip; Step 2: Based on the imaging geometry during the high-speed movement of cold-rolled copper strip, construct a motion blur image synthesis model Mblur_M, and use the motion blur image synthesis model Mblur_M to generate motion blur images corresponding to sharp images, thus constructing a sharp-blur pairing dataset; Step 3: Construct an intelligent de-motion blurring network for cold-rolled copper strip. The intelligent de-motion blurring network includes a feature embedding module, an encoding module, a learnable wavelet fusion module, a wavelet head decoding module, and a multi-scale image output module. Step 4: Train, validate, and test the intelligent motion fuzzing network based on the clear-fuzzy pairing dataset to obtain the intelligent motion fuzzing model DB_M for cold-rolled copper strip; Step 5: Embed the intelligent de-motion blur model DB_M of the cold-rolled copper strip into the automatic visual inspection system for cold rolling production, and restore the motion blur image of the collected cold-rolled copper strip in real time or near real time.

2. The method for synthesizing and restoring motion-blurred images of cold-rolled copper strip according to claim 1, characterized in that: Step 1 includes the following steps: Step 11: Set up an industrial camera at the cold-rolled copper strip production site. The industrial camera is installed on the exit side of the rolling mill or obliquely above the plate roll to collect production images including the surface of the copper strip, the edge contour, or the plate roll area. Step 12: Use a computer to crop the acquired video stream or continuous images, set the image pixel size to H×W, and crop a total of n images as the raw image dataset Raw_dataset; Step 13: Select images from Raw_dataset that are clear, have intact copper band regions, and have no obvious exposure anomalies to construct the clear source image dataset RawClear_dataset.

3. The method for synthesizing and restoring motion-blurred images of cold-rolled copper strip according to claim 1, characterized in that: In step 2, the motion blur image synthesis model Mblur_M determines the motion blur scale as follows: Let the rolling speed be... The camera exposure time is The lens focal length is The distance from the target to the camera is The pixel size is The length of the motion blur kernel is Under the condition that the direction of copper strip motion is approximately parallel to the camera image plane, the length of the motion blur kernel... satisfy: ;in, Used to characterize the pixel displacement on the image plane caused by the movement of the copper strip during camera exposure time; when Using m / min, Using ms, and Using mm, When using μm, the length of the motion blur kernel is expressed as: .

4. The method for synthesizing and restoring motion-blurred images of cold-rolled copper strip according to claim 3, characterized in that: The motion-blurred image synthesis model Mblur_M includes the following steps: Step 21: Select any clear image from the RawClear_dataset. And read or set the rolling speed corresponding to the image. Exposure time ,focal length Object distance and pixel size ; Step 22: According to Computing the fundamental fuzzy scale ;when Using m / min, Using ms, and Using mm, When using μm, the basic fuzzy scale Represented as: Step 23: Set the direction of movement and in Introducing random directional bias on the basis The direction of the fuzzy kernel is obtained. ; Step 24: At the basic fuzzy scale Introducing random scale perturbation The final fuzzy scale is obtained. ; Step 25: According to and Generate motion blur kernel ; Step 26: Clear the image With motion fuzzy kernel Convolution is performed, and random noise used to simulate mill vibration or transverse perturbations in the strip is superimposed. To obtain a synthetic motion-blurred image ,in, This represents the convolution operation. This indicates random noise or spatial disturbance noise.

5. The method for synthesizing and restoring motion-blurred images of cold-rolled copper strip according to claim 4, characterized in that: The random directional deviation The value is a random value within a preset angle range, satisfying: The random scale perturbation For a random value within a preset pixel range, in a preferred embodiment, the following is satisfied: The random noise Transverse random noise or spatial disturbance noise is used to simulate non-ideal imaging factors caused by mill vibration, strip deviation, and micro-vibration of the acquisition system in the cold rolling field.

6. The method for synthesizing and restoring motion-blurred images of cold-rolled copper strip according to claim 1, characterized in that: Step 2 also includes adjusting the fuzzy scale. or rolling speed The synthesized motion-blurred image is divided into multiple blur levels, which include at least slight blur. Moderate blur and severe blur Based on fuzzy scale The fuzzy level classification satisfies: , , ;in, and These are the thresholds between mild and moderate blur, and between moderate and severe blur, respectively; in a preferred embodiment, the threshold is: , Correspondingly, based on rolling speed The fuzzy level classification is as follows: , , .

7. The method for synthesizing and restoring motion-blurred images of cold-rolled copper strip according to claim 1, characterized in that: The intelligent motion deblurring network in step 3 includes: The feature embedding module is used to map the input motion-blurred image into shallow features; The encoding module, consisting of multiple simple encoding blocks (SEBs), is used for progressive downsampling and extraction of multi-scale features; the learnable wavelet fusion module (WFB) is used to transform the encoded features of adjacent scales to the wavelet domain and perform cross-scale fusion; the wavelet head decoding module (WHB) is used to progressively upsample the fused features and reconstruct the image; and the multi-scale output module is used to output the restored images at different scales and participate in multi-scale supervision during the training phase.

8. The method for synthesizing and restoring motion-blurred images of cold-rolled copper strip according to claim 7, characterized in that: Both the learnable wavelet fusion module (WFB) and the wavelet head decoding module (WHB) include learnable wavelet nodes (LWNs), which include forward wavelet convolution, deep feature transformation, and backward wavelet convolution. The forward wavelet convolution is performed by a one-dimensional low-frequency filter. and one-dimensional high-frequency filters Constructing a two-dimensional wavelet convolution kernel using outer product: , , , ,in, For low-frequency subband convolution kernels, For horizontal high-frequency subband convolution kernels, It is a vertical high-frequency subband convolution kernel. The four two-dimensional wavelet convolution kernels are concatenated to form a diagonal high-frequency subband convolution kernel. The complete wavelet convolution kernel is obtained by concatenating these four kernels. For the input feature map It is mapped to the wavelet domain through grouped convolution: And four types of sub-band features were obtained: The one-dimensional low-frequency filter and one-dimensional high-frequency filters The parameters are learnable and can be adaptively updated based on the image features of cold-rolled copper strip during network training; a spatial attention mechanism guided by high-frequency subbands is set in the learnable wavelet node LWN, and the spatial attention mechanism is based on the horizontal high-frequency subbands. and vertical high-frequency sub-band Generate spatial attention weights. First, concatenate the horizontal and vertical high-frequency subbands: Then, spatial attention weights are generated using convolutional mapping and the sigmoid function: ,in, This represents the Sigmoid activation function. Indicates the kernel size as The convolution operation; the spatial attention weights After size alignment, the spatial domain features applied to the wavelet domain features are used to obtain enhanced features: ,in, This indicates element-wise multiplication; through the above method, the network focuses on the edge contour, surface texture, and defect boundary region of cold-rolled copper strip; the depth feature transformation in the learnable wavelet node LWN adopts a lightweight structure, which replaces the 7×7 depth separable convolution with a 3×3 depth separable convolution, and sets 1×1 convolution, GELU activation function, PixelShuffle upsampling branch or element-wise multiplication fusion branch to reduce the number of model parameters and inference time.

9. The method for synthesizing and restoring motion-blurred images of cold-rolled copper strip according to claim 1, characterized in that: Step 4 includes the following steps: Step 41: Set the model training parameters, which include input image size, batch size, optimizer, initial learning rate, learning rate decay strategy, maximum number of iterations, and weight saving interval; Step 42: Input the motion-blurred images in the training set into the intelligent de-motion blurring network, and output multi-scale restored images; Step 43: According to the first Restored images at various scales With corresponding clear label images Calculate the multi-scale restoration loss: ,in, Indicates the number of output scales. Indicates the first Loss weights at each scale, The pixel-level restoration loss function is defined as follows: ; Step 44: Calculate the wavelet constraint loss based on the perfect reconstruction constraints of the wavelet filter. Let the forward wavelet filter be... , The inverse wavelet filter is , Then the perfect reconstruction constraint is satisfied: The corresponding wavelet constraint loss is: ; Step 45: Update the network parameters according to the total loss function, which is: ,in, The weights are set for wavelet-constrained loss. Step 46: Select the model weights with the best overall performance on the validation set as the intelligent motion fuzzing model DB_M for cold-rolled copper strip.

10. The method for synthesizing and restoring motion-blurred images of cold-rolled copper strip according to claim 1, characterized in that: Step 5 includes: inputting the real-time image of cold-rolled copper strip captured by an industrial camera into the intelligent motion blur removal model DB_M of cold-rolled copper strip, outputting the restored clear image, and using the clear image as input for strip width measurement, deviation detection, semantic segmentation, surface defect detection, or plate shape visual inspection models; specifically, let the motion blur image captured by the industrial camera in real time be... The mapping function of the intelligent motion-de-fuzzy model DB_M for cold-rolled copper strip is: The restored clear image Represented as: The restored clear image Input downstream automatic visual inspection model The test results were obtained. The detection results This includes at least one of the following: strip edge position, strip width, deviation amount, semantic segmentation result, surface defect detection result, or plate shape visual inspection result.

Images