Methods, optimization methods, and systems for obtaining high-speed low-light image optimization models.

By extracting features through the VGG19 network and training with the WGAN model, a high-speed low-light image optimization model is generated. This solves the problems of low illumination and noise in the deformation measurement of high-speed rotating parts in a vacuum environment under sealed space, and realizes image enhancement and denoising, meeting the accuracy requirements of high-speed image deformation detection.

CN121190358BActive Publication Date: 2026-11-17XI AN JIAOTONG UNIV +1
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202511283088.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2025-09-09
Publication Date
2026-11-17
Estimated Expiration
2045-09-09

AI Technical Summary

Technical Problem

When measuring the deformation of a high-speed rotating component in a vacuum environment under sealed conditions, the images suffer from low illumination and high noise, resulting in low reliability of feature extraction from the deformed images and failing to meet the accuracy requirements of high-speed image deformation detection.

Method used

Image features are extracted using the VGG19 network and iteratively trained using the WGAN model to generate a high-speed low-light image optimization model. The image is then optimized, enhanced, and denoised using histogram equalization.

Benefits of technology

It improves image quality, meets the accuracy requirements of high-speed image deformation detection, effectively removes noise, repairs image detail loss caused by low illumination and noise, and restores the true features of the image.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121190358B_ABST
    Figure CN121190358B_ABST
Patent Text Reader

Abstract

This invention belongs to the field of image processing technology and relates to a method, optimization method, and system for obtaining a high-speed low-light image optimization model. The invention acquires a set of real images of a rotating body, including a clear still image and a corresponding high-speed low-light image. The set of real images is divided into a first dataset and a second dataset. The clear still image in the first dataset undergoes simulated degradation processing to generate a blurred image. A blurred image dataset is constructed based on the blurred image and the high-speed low-light image in the first dataset. A VGG19 network is used to extract image features from the second dataset and the blurred image dataset. The WGAN model is iteratively trained using the image features, the second dataset, and the blurred image dataset to obtain a high-speed low-light image optimization model. This model can remove noise from high-speed low-light images and enhance them.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, and relates to a method for obtaining, optimizing, and system for high-speed low-light image optimization models. Background Technology

[0002] In high-speed image measurement, DIC (Digital Image Correlation) technology has advantages such as non-contact, high precision, and wide application range. However, when measuring the deformation of high-speed rotating parts in a vacuum environment under sealed space, it suffers from problems such as low illumination and high noise, resulting in low reliability of deformation image feature extraction and failing to meet the accuracy requirements of high-speed image deformation detection.

[0003] Specifically, the measurement environment of the high-speed rotating disk is in a vacuum-sealed box. Due to the uneven lighting in the sealed environment, the acquired images suffer from low illumination and high noise. Conventional methods use filtering or deep learning models (CNN) for denoising, which results in the loss of high-frequency information and harsh edges. Using filtering to process all images uniformly cannot guarantee the quality of all images. Processing each image individually is labor-intensive, inefficient, time-consuming, and can only process a single type of noise at a time. Using deep learning models to process images cannot guarantee the high-frequency details of denoising. For high-speed image deformation detection, the denoising effect of deep learning models (CNN) cannot meet the image accuracy requirements. Summary of the Invention

[0004] The purpose of this invention is to provide a method, optimization method and system for obtaining a high-speed low-light image optimization model, so as to solve the technical problem of poor noise reduction effect of high-speed rotating wheel images.

[0005] To achieve the above objectives, the present invention employs the following technical solution: In a first aspect, the present invention provides a method for obtaining a high-speed low-light image optimization model, comprising the following steps: Acquire a set of real images of the rotating body, the set of real images including a clear still image and a corresponding high-speed low-light image; The real image group is divided into a first dataset and a second dataset. The clear still images in the first dataset are subjected to simulated degradation processing to generate blurred images. The blurred images and the high-speed low-light images in the first dataset constitute a blurred image dataset. The VGG19 network was used to extract image features from the second dataset and the blurred image dataset; The WGAN model was iteratively trained using image features, a second dataset, and a blurred image dataset to obtain a high-speed low-light image optimization model.

[0006] Secondly, this invention provides a high-speed low-light image optimization method, based on a high-speed low-light image optimization model, comprising: Obtain the high-speed low-light image to be optimized; The histogram equalization method and the high-speed low-light image optimization model are used to optimize the high-speed low-light image to obtain the optimized image.

[0007] Thirdly, the present invention provides a system for obtaining a high-speed low-light image optimization model, comprising: The data acquisition module is used to acquire a set of real images of the rotating body, the set of real images including a clear still image and a corresponding high-speed low-light image; The data preprocessing module is used to divide the real image group into a first dataset and a second dataset, perform simulated degradation processing on the clear still images in the first dataset to generate blurred images, and construct a blurred image dataset based on the blurred images and the high-speed low-light images in the first dataset. The image feature acquisition module is used to extract image features from the second dataset and the blurred image dataset using the VGG19 network. The model training module is used to iteratively train the WGAN model using image features, a second dataset, and a blurred image dataset to obtain a high-speed low-light image optimization model.

[0008] Compared with the prior art, the present invention has the following beneficial effects: This invention acquires a set of real images of rotating bodies to provide foundational data for subsequent model training. The clear, still images in the real image set serve as ideal references, while the high-speed, low-light images reflect the low illumination and noise issues encountered in actual measurements, providing raw materials for simulating real measurement scenarios and training the model. The real image set is divided into a first dataset and a second dataset. The clear, still images in the first dataset undergo simulated degradation processing to generate blurred images, enriching the dataset's diversity. The blurred images, along with the high-speed, low-light images in the first dataset, constitute a blurred image dataset, simulating images with different degrees of degradation. This allows the model to learn more features to cope with image degradation during training, improving its ability to handle various complex situations. The VGG19 network is used to extract image features from the second dataset and the blurred image dataset. Extracting image features through the VGG19 network allows for the acquisition of deep semantic information and structural features in the images, helping the WGAN model better learn the mapping relationships between images and improving the model's training effect and performance. The WGAN model is iteratively trained using image features, the second dataset, and the blurred image dataset to obtain an optimized model for high-speed, low-light images. The high-speed low-light image optimization model, after thorough training, has the ability to enhance and denoise high-speed low-light images, effectively improving image quality.

[0009] This invention acquires high-speed low-light images to be optimized, providing raw material for subsequent optimization processing. Histogram equalization and a high-speed low-light image optimization model are used to optimize the images, resulting in optimized images. Histogram equalization redistributes the grayscale values ​​of image pixels, expanding the dynamic range of the image and making the grayscale distribution more uniform. The high-speed low-light image optimization model is trained on a large amount of prior data and possesses the ability to deeply process the low-light and noise problems unique to high-speed low-light images. It can remove noise from the image, repair the loss of image details caused by low light and noise, accurately restore the true features of the image, and make the image clearer and more accurate, meeting the image accuracy requirements of high-speed image deformation detection.

[0010] The system of this invention includes: a data acquisition module, a data preprocessing module, an image feature acquisition module, and a model training module. The data acquisition module acquires a set of real images of the rotating body, including a clear still image and a corresponding high-speed low-light image. The data preprocessing module divides the real image set into a first dataset and a second dataset, performs simulated degradation processing on the clear still image in the first dataset to generate a blurred image, and constructs a blurred image dataset based on the blurred image and the high-speed low-light image in the first dataset. The image feature acquisition module uses a VGG19 network to extract image features from the second dataset and the blurred image dataset. The model training module iteratively trains a WGAN model using the image features, the second dataset, and the blurred image dataset to obtain an optimized model for high-speed low-light images. These modules work together to enhance and denoise high-speed low-light images, effectively improving image quality. Attached Figure Description

[0011] Figure 1 This is a flowchart of the method for obtaining a high-speed low-light image optimization model according to an embodiment of the present invention; Figure 2 This is a flowchart of the high-speed low-light image optimization method according to an embodiment of the present invention; Figure 3 This is a system module connection diagram according to an embodiment of the present invention; Figure 4 This is a flowchart of the high-speed low-light image enhancement and denoising method based on the WGAN model according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the structural configuration of the high-speed rotating component image acquisition system according to an embodiment of the present invention; Figure 6 This is a network structure diagram of the WGAN model discriminator module according to an embodiment of the present invention; Figure 7This is a flowchart illustrating the overall algorithm of the high-speed low-light image enhancement and denoising method based on the WGAN model in this embodiment of the invention.

[0012] The components include: 1. Binocular measurement unit; 2. Test sealed box; 3. Rotating body; 4. High-intensity light source; 5. Synchronous control box; 6. Observation window; 11. Left high-speed camera; 12. Right high-speed camera. Detailed Implementation

[0013] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0014] It should be noted that the terms "first," "second," etc., in the specification and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0015] The present invention will now be described in further detail with reference to the accompanying drawings: See Figure 1 This invention discloses a method for obtaining a high-speed low-light image optimization model, comprising the following steps: S1, Obtain a set of real images of the rotating body. The set of real images includes a clear still image and a corresponding high-speed low-light image, which provides basic data for subsequent model training. The clear still image in the set of real images can be used as an ideal reference, while the high-speed low-light image reflects the image conditions such as low illumination and noise encountered in actual measurement, providing original materials for simulating real measurement scenarios and training models.

[0016] In this embodiment of the invention, obtaining the real image set of the rotating body includes: The system acquires real images of a rotating body under different rotation speeds and light intensities from a high-speed image acquisition system. The real image acquired when the rotation speed of the rotating body is zero is a clear, static image, and the real image acquired when the rotation speed of the rotating body is not zero is a high-speed, low-light image.

[0017] In this embodiment of the invention, the real image acquired when the rotational speed of the rotating body is zero is a clear still image, and the step of acquiring the clear still image further includes: For each acquisition process, it is necessary to acquire long-exposure images at low ISO and short-exposure images at high ISO. For the same acquisition process, multiple clear still images are captured consecutively, aligned, weighted averaged and fused, and random noise is reduced.

[0018] S2, the real image set is divided into a first dataset and a second dataset. The clear still images in the first dataset undergo simulated degradation processing to generate blurred images, enriching the diversity of the dataset. The blurred images, together with the high-speed low-light images in the first dataset, constitute a blurred image dataset, capable of simulating images with different degrees of degradation. This allows the model to learn more features to cope with image degradation during training, improving its ability to handle various complex situations. Simultaneously, two datasets are divided: one part is used to generate the blurred image dataset to assist training, and the other part (the second dataset) retains the original real image features for subsequent joint training of the model with the blurred image dataset, ensuring the comprehensiveness and rationality of the training data.

[0019] In this embodiment of the invention, the step of performing simulated degradation processing on the clear still images in the first dataset to generate blurred images includes: The clear still images in the first dataset are translated, stretched, and rotated to deform them, and the pixel brightness is reduced by linear scaling to generate images with different dark light levels. Gaussian noise is added to the generated images at different levels of darkness to obtain blurred images; The method of reducing pixel brightness through linear scaling is specifically formulated as follows:

[0020] In the formula: The original image matrix, This is the original image matrix after brightness compression. The compression factor, ∈(0,1); The Gaussian noise is added to the generated images at different levels of darkness, using the following formula:

[0021] In the formula: This is the original image matrix after adding Gaussian noise. The noise follows a normal Gaussian distribution. Let Variance be the variance.

[0022] S3 employs the VGG19 network (Visual Geometry Group 19-layer Network) to extract image features from the second dataset and the fuzzy image dataset. By extracting image features through the VGG19 network, deep semantic information and structural features in the images can be obtained, which helps the WGAN model to better learn the mapping relationship between images and improve the training effect and performance of the model.

[0023] S4 uses image features, a second dataset, and a blurred image dataset to iteratively train the WGAN model (Wasserstein Generative Adversarial Network) to obtain a high-speed low-light image optimization model. The resulting high-speed low-light image optimization model, after thorough training, possesses the ability to enhance and denoise high-speed low-light images, effectively improving image quality and providing clear and reliable image data for subsequent high-speed image deformation detection.

[0024] In this embodiment of the invention, the step of iteratively training the WGAN model using image features, a second dataset, and a blurred image dataset to obtain a high-speed low-light image optimization model includes: The image features, the second dataset, and the blurred image dataset are input into the WGAN model, and the loss during the iteration cycle is obtained as follows: Input the high-speed low-light images from the second dataset or the blurred images from the blurred image dataset into the generator. After the generator calculates, it generates the generator image. The generator image is input into the discriminator, and a clear still image is used as the ground truth image and input into the discriminator. The discriminator calculates and judges the result to obtain the discriminator's loss. The formula for obtaining the loss of the discriminator is as follows:

[0025] In the formula, For the loss of the discriminator, It is the first The result of forward propagation of a blurred image through a discriminator It is the first The result of passing the true value image through the discriminator It is the W distance between the ground truth image and the blurred image. It is the first Zhang's truth image distribution, They are the first Zhang's generated image distribution; The blurred image is forward-propagated in the discriminator, and the average value of the discriminator output is used as the generator loss. The generator loss is obtained by linearly superimposing the perceptual loss, pixel loss and generation loss after weighting. The formula for obtaining the generator loss is as follows:

[0026]

[0027] In the formula: It is the first Zhang's true value image features, It is the first Zhang's blurred image features It is the first The result of forward propagation of the generated image through the discriminator. It is the generator loss. This is the weighting coefficient for perceptual loss, set to 10 in the model. These are the weighting coefficients for pixel loss, set to 10000 in the model. It is perceived loss. It's a pixel loss. It is the implicit generator loss. The number of true images.

[0028] The loss within the iteration cycle is backpropagated to obtain the gradient and loss of each neuron in the layer. The weights of each layer are updated based on the gradient and loss of each neuron to reduce the loss within the iteration cycle. The WGAN model is iteratively trained to obtain a high-speed low-light image optimization model, including: Based on the loss within the iteration cycle, the gradient and loss of each neuron in each layer are calculated in reverse layer by layer using the chain rule; The parameters of the WGAN model are updated based on the gradient and loss of each neuron in each layer, thereby reducing the loss within the iteration cycle. The parameter update formula is as follows:

[0029]

[0030] In the formula: It is the initial learning rate; It is the new learning rate; It is a coefficient used to control the decay of the first-order gradient moment estimate, and it is between 0 and 1; It is a coefficient used to control the decay of the second-order gradient moment estimate, and it is between 0 and 1; It is the result of the first moment estimation during the iteration process; It is the result of the second moment estimation during the iteration process; These are the model parameters before the update; These are the updated model parameters; This is the error term. After obtaining the gradients of each layer through backpropagation, the first and second moments of the gradients are estimated and used to update the learning rate. The updated learning rate is then used for training in the next cycle.

[0031] In this embodiment of the invention, the following steps are also included: A clear still image and a corresponding high-speed low-light image are input into a high-speed low-light image optimization model to obtain a denoised image. The denoising performance of the model was evaluated by comparing denoised images, high-speed low-light images, and sharp still images, based on PSNR and SSIM.

[0032] Among them, PSNR (Peak Signal-to-Noise Ratio) is used to measure the difference between two images; SSIM (Structural Similarity) can assess image similarity in a way similar to human perception.

[0033] This invention provides a high-speed low-light image enhancement and denoising method based on the WGAN model, which comprehensively solves the problem of low reliability in feature extraction of deformed images due to low illumination and high noise levels when measuring the deformation of high-speed rotating parts in a vacuum environment under sealed conditions, thus failing to meet the accuracy requirements of high-speed image deformation detection. By using the WGAN model to fuse the acquired images, a ground truth image is obtained, reducing the impact of spatial misalignment on the image and improving image accuracy. Image data augmentation is used to expand the dataset, saving time and cost in dataset construction. Image feature extraction is performed using a VGG19 pre-trained model, providing high-quality feature input for model training. Through multiple iterations of forward and backward propagation to update model parameters, combined with histogram equalization, effective enhancement and denoising of high-speed low-light images are finally achieved, resulting in clear, high-contrast speckle images that meet the accuracy requirements of high-speed image deformation detection. This provides a reliable and efficient image processing method for the field of high-speed image measurement.

[0034] See Figure 2 The present invention also discloses a high-speed low-light image optimization method, based on a high-speed low-light image optimization model, comprising: S1: Obtain the high-speed low-light image to be optimized, providing raw material for subsequent optimization processing.

[0035] S2 employs histogram equalization and a high-speed low-light image optimization model to optimize the high-speed low-light image to obtain the optimized image. Histogram equalization expands the dynamic range of the image by redistributing the gray values ​​of image pixels, making the gray-level distribution of the image more uniform. The high-speed low-light image optimization model is trained on a large amount of prior data and has the ability to deeply process the low-light and noise problems unique to high-speed low-light images. It can learn the distribution pattern of noise in the image and the mapping relationship between clear images and low-light noisy images. Based on histogram equalization processing, it further removes noise from the image, repairs the loss of image details caused by low light and noise, accurately restores the true features of the image, and makes the image clearer and more accurate, meeting the image accuracy requirements of high-speed image deformation detection.

[0036] See Figure 3 Based on the above method, this invention also discloses a system for obtaining a high-speed low-light image optimization model, comprising: The data acquisition module is used to acquire a set of real images of the rotating body, the set of real images including a clear still image and a corresponding high-speed low-light image; The data preprocessing module is used to divide the real image group into a first dataset and a second dataset, perform simulated degradation processing on the clear still images in the first dataset to generate blurred images, and construct a blurred image dataset based on the blurred images and the high-speed low-light images in the first dataset. The image feature acquisition module is used to extract image features from the second dataset and the blurred image dataset using the VGG19 network. The model training module is used to iteratively train the WGAN model using image features, a second dataset, and a blurred image dataset to obtain a high-speed low-light image optimization model.

[0037] The various modules of the system in this invention work together to enhance and denoise high-speed low-light images, effectively improving image quality.

[0038] Example 2: See Figure 4 and Figure 6 This embodiment proposes a high-speed low-light image optimization method based on the WGAN model, which can meet the requirements of image enhancement and denoising for high-speed low-light images, including the following steps: S1: A high-speed image acquisition system is built using two high-speed cameras and a strong light source to acquire real images, including clear still images and high-speed low-light images. Preferably, step S1 specifically includes: S11: A high-speed image acquisition system consisting of two high-speed cameras and a high-intensity light source; S12: Fix the positions and light intensity of two high-speed cameras, control the rotation speed of the rotating body, and synchronously acquire clear still images of the rotating body and high-speed low-light images at different rotation speeds; Preferably, step S12 specifically includes: To ensure that the acquired still images are clear and low in noise, the image acquisition process and the acquired images need to be processed. For each acquisition process, it is necessary to acquire long-exposure images at low ISO (International Organization for Standardization) and short-exposure images at high ISO, while avoiding camera shake, minimizing spatial misalignment, aligning multiple clear still images captured consecutively in the same acquisition process, weighted averaging and fusing them, and reducing random noise.

[0039] High sensitivity is ISO 6400, short exposure is exposure time of 0.01s, low sensitivity is ISO 100, and long exposure is exposure time of 5s.

[0040] S13: Adjust the light source intensity, control the rotation speed of the rotating body, and simultaneously acquire clear still images of different dark light levels when the rotating body is stationary and high-speed dark light images of different dark light levels at different rotation speeds.

[0041] S2: Construct an image dataset based on real images, divide the image dataset into a training set and a validation set, deform some clear still images, add random noise and low brightness processing to simulate degraded images, obtain blurred images, and add the blurred images and corresponding clear still images to the training set. Preferably, step S2 specifically includes: S21: Construct an image dataset and divide it into a training set and a validation set. Use a small portion of real images as the validation set and the other collected images as the training set. Use a portion of clear and low-noise still images in the training set as real images and simulate degradation of the remaining clear still images, as detailed in S22. S22: Transform a clear still image by translating, stretching, rotating, etc., add random noise and perform low brightness processing to simulate the generation of a blurred image, and add it to the training set; Preferably, step S22 specifically includes: The image is transformed by algorithms through translation, stretching, rotation, etc., and then the brightness is reduced to generate images with different levels of darkness. Gaussian noise is added to simulate interference noise and perform image data enhancement. Add pixels to the image with a mean of 0 and a variance of 0. Gaussian noise:

[0042] In the formula: The original image matrix, This is the original image matrix after adding Gaussian noise. The noise follows a normal Gaussian distribution. Reduce pixel brightness (compression factor) through linear scaling. ∈(0,1)):

[0043] In the formula: The original image matrix, This is the original image matrix after brightness compression. This is the compression factor.

[0044] S3: Pre-train the VGG19 model, extract images, start WGAN model training, and read the degraded image into the model in each training iteration to calculate the loss within the iteration cycle; Preferably, step S3 specifically includes: S31: By using a pre-trained VGG19 network to extract image features from generated and real images in the training set, the VGG19 network has been trained on large-scale datasets and is widely used for high-level image feature extraction. S32: Start WGAN model training. In each training iteration, read the blurred image into the model and perform forward propagation to calculate the loss within the iteration.

[0045] Preferably, step S32 specifically includes: In each training iteration, the image generated by the generator is imported into the discriminator for forward propagation to obtain the loss within the iteration period; The loss within the generator's iteration cycle is obtained by linearly summing the weighted perceptual loss, pixel loss, and generation loss, as shown in the following formula:

[0046]

[0047] In the formula: These are real image features. It generates image features. It is the result of the generated image being forward-propagated through the discriminator. These are the distributions of the real images. It generates image distribution. It is the complete generator loss. These are the weighting coefficients for the perceptual loss, set to 10 in the model. These are the weighting coefficients for pixel loss, set to 10000 in the model. It is perceived loss. It's a pixel loss. It is the implicit generator loss. The number of true images.

[0048] S4: Backpropagate the loss within the iteration cycle to calculate the loss of each neuron in the layer, update the weights of each neuron in the layer to reduce the loss, and continue iterating until the WGAN model training is complete. Preferably, step S4 specifically includes: After obtaining the loss within the iteration cycle through forward propagation, the loss within the iteration cycle is backpropagated to calculate the loss of each neuron in the layer. Based on the loss of each neuron in the layer, the weights of each neuron in the layer are updated using the gradient descent method, so that the loss of the model during forward propagation after updating the parameters is reduced. The Adam optimization algorithm is used to calculate the model parameters. The formula for parameter update is as follows:

[0049]

[0050] In the formula: It is the initial learning rate. It's the new learning rate. This is a coefficient used to control the decay of the first-order gradient moment estimate; its value is between 0 and 1. This is a coefficient used to control the decay of the second-order gradient moment estimate; its value is between 0 and 1. It is the first moment estimation result during the iteration process. It is the second-order moment estimation result during the iteration process. and These are the model parameters before and after the update. This is the error term. After obtaining the gradients of each layer through backpropagation, the first and second moments of the gradients are estimated and used to update the learning rate. The updated learning rate is then used for training in the next cycle.

[0051] S5: Histogram equalization is used to enhance the high-speed low-light image to be optimized, and the image denoising effect of the model is verified using a validation set.

[0052] Preferably, step S5 specifically includes: S51: Histogram equalization is used to enhance high-speed low-light images; S52: The model's image denoising performance is validated using a validation set. Real images from the validation set are input into the model for forward propagation, yielding the denoised image generated by the generator. The denoised image, the real image, and the denoised image are compared to evaluate the model's denoising performance. This is primarily evaluated using PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity). PSNR measures the difference between two images, while SSIM assesses image similarity in a manner similar to human perception.

[0053] Another embodiment of the present invention discloses a computer-readable storage medium storing computer-executable instructions. When the computer-executable instructions are invoked and executed by a processor, the computer-executable instructions cause the processor to implement the steps of the above-described high-speed low-light image enhancement and denoising method.

[0054] In summary, compared with existing technologies, the high-speed low-light image enhancement and denoising method based on the WGAN model provided by this invention utilizes the WGAN model to obtain a ground truth image through the fusion of acquired images, reducing the impact of spatial misalignment. Image data augmentation is used to expand the dataset, saving time in dataset construction. Image extraction is performed using a VGG19 pre-trained model, and model parameters are continuously updated through multiple iterations of forward and backward propagation. Combined with histogram equalization, this method enhances and denoises high-speed low-light images, resulting in clear, high-contrast speckle images that meet the accuracy requirements of high-speed image deformation detection.

[0055] Example 3: Existing high-speed low-light image enhancement and denoising schemes generally use Gaussian filtering or CNN for denoising, which suffers from problems such as loss of high-frequency information and harsh edges. Therefore, this invention provides a new method for high-speed low-light image enhancement and denoising based on the WGAN model.

[0056] like Figure 1 As shown, one embodiment of the present invention discloses a high-speed low-light image optimization method, comprising the following five parts: S1: A high-speed image acquisition system is built using two high-speed cameras and a strong light source 4 to acquire clear still images and high-speed low-light images. In a specific embodiment, high-speed cameras and light sources can be used to simulate different environments for image acquisition. S2: Construct an image dataset and divide it into a training set and a validation set. Deform the clear still image, add random noise and low brightness processing to simulate the degraded image and generate a blurred image, which is then added to the training set. In a specific implementation, the training set can be expanded by simulating the synthesis of clear still images using an algorithm. S3: Pre-train the VGG19 model, extract image features, and start training the WGAN model. In each training iteration, a batch of data is read into the model for forward propagation to calculate the loss within the iteration cycle. In a specific implementation, the VGG19 model can be used for image extraction, and then the WGAN model can be trained. S4: Backpropagate the loss within the iteration cycle to calculate the loss of each neuron in the layer, update the weights of each neuron in the layer to reduce the loss, and continue iterating until the WGAN model training is completed. Iterate continuously to optimize the model. S5: Histogram equalization is used to enhance the high-speed low-light image to be optimized. The image denoising effect of the model is verified using a validation set. In a specific implementation, the PSNR and SSIM of the denoised image and the real image can be compared to verify the denoising effect of the model.

[0057] In a further embodiment, a method is proposed to obtain low-noise and spatially misaligned original images and algorithmically simulated degraded images to expand the dataset: long-exposure images are acquired at low ISO and short-exposure images are acquired at high ISO to avoid camera shake; multiple frames of images captured continuously in the same acquisition process are aligned, weighted averaged and fused to synthesize a clear image; and the clear still image is degraded through an algorithm to achieve image data enhancement and expand the dataset.

[0058] In a further embodiment, an image denoising method is provided that can remove various types of noise from high-speed low-light images and more comprehensively preserve image details: First, images are acquired and a dataset is constructed; then, a WGAN model is trained, and the model parameters are iteratively optimized through forward propagation and backward propagation; finally, the denoising effect of the model is verified using denoised images from the validation set and real images.

[0059] In a further embodiment, high-quality original images are acquired, a sufficiently large dataset is constructed, and a WGAN model is trained on this dataset to perform image denoising and low-light enhancement. Specifically, the acquired high-speed low-light images are input into the WGAN model for forward propagation to obtain the denoised images generated by the generator. The high-speed low-light images are then enhanced using histogram equalization to obtain clear, high-contrast speckle images.

[0060] like Figures 4 to 7 As shown, a preferred embodiment of the present invention provides a high-speed low-light image optimization method, comprising the following steps: S1: A high-speed image acquisition system is built using two high-speed cameras and a strong light source to acquire original, clear still images and high-speed images. The specific steps are as follows: S11: A high-speed rotating part image acquisition system is built using a high-speed camera, including a binocular measurement unit 1, which includes a left high-speed camera 11 and a right high-speed camera 12. like Figure 5 As shown, two observation windows 6 are opened on the surface of the test sealed box 2. The left high-speed camera 11 and the right high-speed camera 12 each use one observation window 6 to measure the deformation of the rotating body 3 inside the box. The left high-speed camera 11 and the right high-speed camera 12 are controlled by the synchronous control box 5 to capture the deformation image of the tested object in real time. Since the exposure time is extremely short, a light source 4 is used for supplementary lighting.

[0061] S12: Fix the positions and light source intensity of the two high-speed cameras, control the rotation speed of the rotating body 3, and synchronously acquire still images and high-speed images of the rotating body 3 at different rotation speeds; In order to ensure that the original images acquired are clear and low in noise, it is necessary to process both the image acquisition process and the acquired images. For each acquisition process, it is necessary to acquire long-exposure images at low ISO and short-exposure images at high ISO, while avoiding camera shake, minimizing spatial misalignment, aligning multiple frames of images captured consecutively in the same acquisition process, weighted averaging and fusing them, and reducing random noise.

[0062] S13: Adjust the intensity of the light source 4, control the rotation speed of the rotating body 3, and simultaneously acquire clear still images of different dark light levels when the rotating body 3 is stationary and high-speed dark light images of different dark light levels at different rotation speeds.

[0063] S2: Construct an image dataset, dividing it into a training set and a validation set. The original clear images are deformed, random noise is added, and low-brightness processing is applied to simulate degraded images, generating blurred images which are then added to the training set. The specific steps are as follows: S21: Construct an image dataset and divide it into a training set and a validation set. Use a small portion of the collected images as the validation set and the other collected images as the training set. Use clear and low-noise original images as real images. S22: Transform the synthesized clear image by translating, stretching, rotating, etc., add random noise and perform low brightness processing to simulate the generation of a blurred image, and add it to the training set; The algorithm transforms the image by translating, stretching, and rotating it, then reduces its brightness to generate images with different levels of darkness, and adds Gaussian noise to simulate interference noise for image data enhancement. Add pixels to the image with a mean of 0 and a variance of 0. Gaussian noise: (1) In Equation 1: This is the original image matrix after adding Gaussian noise. The noise follows a normal Gaussian distribution. For variance; Reduce pixel brightness (compression factor) through linear scaling. ∈(0,1)): (2) In Equation 2: The original image matrix, This is the original image matrix after brightness compression. The compression factor, ∈(0,1).

[0064] S3: Pre-train the VGG19 model, extract images, and begin WGAN model training. In each training iteration, a batch of data is read into the model for forward propagation to calculate the loss. The specific steps are as follows: S31: By using a pre-trained VGG19 network to extract image features from generated and real images, the VGG19 network has been trained on large-scale datasets and is widely used for high-level image feature extraction. S32: Start training the WGAN model. In each training iteration, a batch of data is read into the model for forward propagation to calculate the loss.

[0065] In each training iteration, a batch of data is read into the model using a data loader for forward computation. The image generated by the generator is imported into the discriminator for forward propagation to obtain the loss within the iteration cycle, which consists of three parts: perceptual loss, pixel loss, and generation loss. As the number of iterations increases, the overall loss gradually decreases and tends to stabilize. The complete loss of the generator is obtained by linearly summing the weighted perceptual loss, pixel loss, and generation loss, as shown in the following formula:

[0066] (3) In Equation 3: It is the first Zhang's true value image features, It is the first Zhang's blurred image features It is the first The result of forward propagation of the generated image through the discriminator. It is the generator loss. This is the weighting coefficient for perceptual loss, set to 10 in the model. These are the weighting coefficients for pixel loss, set to 10000 in the model. It is perceived loss. It's a pixel loss. It is the implicit generator loss. The number of true images.

[0067] like Figure 6 As shown, the input image sequence is generated by the generator. A combination of global and local discriminators is used to discriminate the generated images. In the global discriminator, the generated image passes through a series of two-dimensional convolutional layers, normalization layers, and ReLU activation layers to continuously reduce the dimensionality and obtain the global output. In the local discriminator, a region is randomly cropped from the generated image as its input. This region also passes through a series of convolutional, normalization, and activation layers to obtain the local output. The global and local outputs are concatenated to form the overall output, which represents the intensity of potential noise in the image.

[0068] S4: Backpropagate the loss to calculate the loss of each neuron in each layer, update the weights of each neuron to reduce the loss, and continue iterating until the WGAN model training is complete. The specific steps are as follows: In this process, after obtaining the loss of the iteration cycle through forward propagation, the loss is backpropagated to calculate the loss of each neuron in the layer. Then, the weights of each neuron in the layer are updated according to the gradient descent method, so that the loss of the model during forward propagation is reduced after updating the parameters. The learning rate needs to be updated after each iteration. If the loss within the cycle is less than the minimum loss, the minimum loss also needs to be updated, and the model parameters are recorded. The Adam optimization algorithm is used to calculate model parameters. This method dynamically adjusts the learning rate by estimating the first and second moments of the gradient, ensuring relatively smooth gradient updates. It avoids the problems of poor convergence caused by an excessively large learning rate and slow training and local convergence caused by an excessively small learning rate. The formula for parameter update is as follows: (4) (5) In Equations 4 and 5: It is the initial learning rate; It is the new learning rate; It is a coefficient used to control the decay of the first-order gradient moment estimate, and it is between 0 and 1; It is a coefficient used to control the decay of the second-order gradient moment estimate, and it is between 0 and 1; It is the result of the first moment estimation during the iteration process; It is the result of the second moment estimation during the iteration process; These are the model parameters before the update; These are the updated model parameters; This is the error term. After obtaining the gradients of each layer through backpropagation, the first and second moments of the gradients are estimated and used to update the learning rate. The updated learning rate is then used for training in the next cycle.

[0069] S5: Histogram equalization is used to enhance high-speed low-light images. The image denoising effect of the model is verified using a validation set. The specific steps are as follows: S51: Histogram equalization is used to enhance low-light images; S52: The model's image denoising performance is validated using a validation set. The acquired images from the validation set are input into the model for forward propagation to obtain the denoised image generated by the generator. The denoised image, the acquired image, and the ground truth image are compared to evaluate the model's denoising performance, primarily using PSNR and SSIM for image evaluation.

[0070] This invention first constructs a high-speed image acquisition system to acquire original clear images and high-speed low-light images. Second, it builds an image dataset, dividing it into a training set and a validation set. Then, in each training iteration, a batch of data is read into the model for forward propagation to calculate the loss. Next, this loss is backpropagated to calculate the loss of each neuron in each layer, and the weights of each neuron are updated. Finally, histogram equalization is used to enhance the image, and the validation set is used to verify the model's image denoising performance. This invention can achieve enhancement and denoising of high-speed low-light images, meeting the accuracy requirements for deformation detection in high-speed low-light images.

[0071] This invention effectively solves the problem of low illumination in high-speed low-light images. The solution optimizes only the algorithm level and does not require additional equipment to enhance contrast, thus effectively controlling cost and system complexity. The image noise removal method proposed in this invention achieves various noise removals in high-speed low-light images by constructing a dataset and continuously iteratively optimizing the WGAN model, while preserving high-frequency edge information. Combined with histogram equalization, it can obtain clear, high-contrast speckle images that meet the accuracy requirements for deformed image measurement.

[0072] The above content is only for illustrating the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solution based on the technical concept proposed in this invention shall fall within the scope of protection of this invention.

Claims

1. A method for obtaining a high-speed low-light image optimization model, characterized in that, Includes the following steps: Acquire a set of real images of the rotating body, the set of real images including a clear still image and a corresponding high-speed low-light image; The real image group is divided into a first dataset and a second dataset. The clear still images in the first dataset are subjected to simulated degradation processing to generate blurred images. The blurred images and the high-speed low-light images in the first dataset constitute a blurred image dataset. The step of simulating degradation processing on the clear still images in the first dataset to generate blurred images includes: The clear still images in the first dataset are translated, stretched, and rotated to deform them, and the pixel brightness is reduced by linear scaling to generate images with different dark light levels. Gaussian noise is added to the generated images at different levels of darkness to obtain blurred images; The method of reducing pixel brightness through linear scaling is specifically formulated as follows: In the formula: The original image matrix, This is the original image matrix after brightness compression. The compression factor, ∈(0,1); The Gaussian noise is added to the generated images at different levels of darkness, using the following formula: In the formula: This is the original image matrix after adding Gaussian noise. The noise follows a normal Gaussian distribution. For variance; The VGG19 network was used to extract image features from the second dataset and the blurred image dataset; The WGAN model was iteratively trained using image features, a second dataset, and a blurred image dataset to obtain a high-speed low-light image optimization model.

2. The method for obtaining a high-speed low-light image optimization model according to claim 1, characterized in that, The acquisition of the real image set of the rotating body includes: The system acquires real images of a rotating body under different rotation speeds and light intensities from a high-speed image acquisition system. The real image acquired when the rotation speed of the rotating body is zero is a clear, static image, and the real image acquired when the rotation speed of the rotating body is not zero is a high-speed, low-light image.

3. The method for obtaining a high-speed low-light image optimization model according to claim 2, characterized in that, The real image acquired when the rotational speed of the rotating body is zero is a clear still image, and the acquisition step of the clear still image further includes: For each acquisition process, it is necessary to acquire long-exposure images at low ISO and short-exposure images at high ISO. For the same acquisition process, multiple clear still images are captured consecutively, aligned, weighted averaged and fused, and random noise is reduced.

4. The method for obtaining the high-speed low-light image optimization model according to claim 1, characterized in that, The method involves iteratively training the WGAN model using image features, a second dataset, and a blurred image dataset to obtain a high-speed low-light image optimization model, including: Input the image features, the second dataset, and the blurred image dataset into the WGAN model and obtain the loss within the iteration cycle; The loss within the iteration cycle is backpropagated to obtain the gradient and loss of each neuron in the layer. The weights of each layer are updated based on the gradient and loss of each neuron in the layer, so that the loss within the iteration cycle is reduced. The WGAN model is iteratively trained to obtain a high-speed low-light image optimization model.

5. The method for obtaining a high-speed low-light image optimization model according to claim 4, characterized in that, The image features, the second dataset, and the blurred image dataset are input into the WGAN model to obtain the loss within the iteration cycle, as detailed below: Input the high-speed low-light images from the second dataset or the blurred images from the blurred image dataset into the generator. After the generator calculates, it generates the generator image. The generator image is input into the discriminator, and a clear still image is used as the ground truth image and input into the discriminator. The discriminator calculates and judges the result to obtain the discriminator's loss. The formula for obtaining the loss of the discriminator is as follows: In the formula, For the loss of the discriminator, It is the first The result of forward propagation of a blurred image through a discriminator It is the first The result of passing the true value image through the discriminator It is the W distance between the ground truth image and the blurred image. It is the first Zhang's truth image distribution, They are the first Zhang's generated image distribution; The blurred image is forward-propagated in the discriminator, and the average value of the discriminator output is used as the generator loss. The generator loss is obtained by linearly superimposing the perceptual loss, pixel loss and generation loss after weighting. The formula for obtaining the generator loss is as follows: In the formula: It is the first Zhang's true value image features, It is the first Zhang's blurred image features It is the first The result of forward propagation of the generated image through the discriminator. It is the generator loss. This is the weighting coefficient for perceptual loss, set to 10 in the model. These are the weighting coefficients for pixel loss, set to 10000 in the model. It is perceived loss. It's a pixel loss. It is the implicit generator loss. The number of true images.

6. The method for obtaining a high-speed low-light image optimization model according to claim 4, characterized in that, The step of backpropagating the loss within the iteration cycle to obtain the gradient and loss of each neuron in each layer, and updating the weights of each layer based on the gradient and loss of each neuron in each layer to reduce the loss within the iteration cycle includes: Based on the loss within the iteration cycle, the gradient and loss of each neuron in each layer are calculated in reverse layer by layer using the chain rule; The parameters of the WGAN model are updated based on the gradient and loss of each neuron in each layer, thereby reducing the loss within the iteration cycle. The parameter update formula is as follows: In the formula: It is the initial learning rate; It is the new learning rate; It is a coefficient used to control the decay of the first-order gradient moment estimate, and it is between 0 and 1; It is a coefficient used to control the decay of the second-order gradient moment estimate, and it is between 0 and 1; It is the result of the first moment estimation during the iteration process; It is the result of the second moment estimation during the iteration process; These are the model parameters before the update; These are the updated model parameters; This is the error term.

7. The method for obtaining a high-speed low-light image optimization model according to claim 1, characterized in that, It also includes the following steps: A clear still image and a corresponding high-speed low-light image are input into a high-speed low-light image optimization model to obtain a denoised image. The denoising performance of the model was evaluated by comparing denoised images, high-speed low-light images, and sharp still images, based on PSNR and SSIM.

8. A high-speed low-light image optimization method, based on the high-speed low-light image optimization model according to any one of claims 1 to 7, characterized in that, include: Obtain the high-speed low-light image to be optimized; The histogram equalization method and the high-speed low-light image optimization model are used to optimize the high-speed low-light image to obtain the optimized image.

9. A system for acquiring a high-speed low-light image optimization model, characterized in that, include: The data acquisition module is used to acquire a set of real images of the rotating body, the set of real images including a clear still image and a corresponding high-speed low-light image; The data preprocessing module is used to divide the real image group into a first dataset and a second dataset, perform simulated degradation processing on the clear still images in the first dataset to generate blurred images, and construct a blurred image dataset based on the blurred images and the high-speed low-light images in the first dataset. The step of simulating degradation processing on the clear still images in the first dataset to generate blurred images includes: The clear still images in the first dataset are translated, stretched, and rotated to deform them, and the pixel brightness is reduced by linear scaling to generate images with different dark light levels. Gaussian noise is added to the generated images at different levels of darkness to obtain blurred images; The method of reducing pixel brightness through linear scaling is specifically formulated as follows: In the formula: The original image matrix, This is the original image matrix after brightness compression. The compression factor, ∈(0,1); The Gaussian noise is added to the generated images at different levels of darkness, using the following formula: In the formula: This is the original image matrix after adding Gaussian noise. The noise follows a normal Gaussian distribution. For variance; The image feature acquisition module is used to extract image features from the second dataset and the blurred image dataset using the VGG19 network. The model training module is used to iteratively train the WGAN model using image features, a second dataset, and a blurred image dataset to obtain a high-speed low-light image optimization model.

Citation Information

Patent Citations

  • Hyperspectral image denoising method based on blind noise estimation

    CN112634159A

  • Weak light image enhancement method and device based on conditional adversarial network

    CN113674191A