Power transmission line image denoising method based on multi-scale EPLL

By employing a multi-scale EPLL image denoising method, combined with semi-quadratic splitting optimization and a multi-scale framework, the likelihood logarithm expectation of image patches is optimized, thus solving the problem of low detection accuracy in transmission line image detection and achieving efficient and accurate transmission line defect identification.

CN119048386APending Publication Date: 2024-11-29GUODIAN NANJING AUTOMATION
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411105610.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-13
Publication Date
2024-11-29

AI Technical Summary

Technical Problem

Existing technologies for power transmission line image inspection suffer from low detection accuracy and high false positive and false negative rates, especially in the poor identification of small defects. Furthermore, the lack of optimization in image denoising algorithms leads to low detection accuracy.

Method used

An image denoising method based on multi-scale EPLL is adopted. By constructing the prior distribution parameters of image patches, combining the semi-quadratic split optimization method and multi-scale frame, a single-pass multi-frame detector is used for preprocessing, and denoising is performed channel by channel in different color spaces. The likelihood log expectation of image patches is optimized to achieve efficient denoising.

Benefits of technology

It improves the detection accuracy and clarity of transmission line images, reduces artifacts, preserves image details and edge information, enhances the ability to identify small defects, and achieves real-time and efficient detection results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119048386A_ABST
    Figure CN119048386A_ABST
Patent Text Reader

Abstract

The invention discloses a multi-scale EPLL-based power transmission line image denoising method, and relates to the field of power system image processing, and the method comprises the following steps: obtaining a power transmission line image, and preprocessing the power transmission line image to obtain an image data set; extracting an image block based on the image data set, and eliminating a direct current component; calculating prior distribution parameters of the image blocks after the direct current component is eliminated based on an expectation maximization algorithm, and checking convergence of likelihood logarithm of the preprocessed power transmission line image; and in combination with a semi-quadratic splitting optimization method, minimizing an energy function, and alternately optimizing the image blocks and the power transmission line image to realize image denoising. According to the method, denoising can be executed through GMM modeling based on the prior of the image blocks, the performance of the algorithm is improved by optimizing the parameters of the image blocks, meanwhile, a multi-scale framework is added to be applied to different single-scale denoising algorithms, the single-scale denoising algorithms are converted to different color spaces for denoising, and the artifact phenomenon is reduced.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of power system image processing, and in particular to a multi-scale EPLL transmission line image denoising method. Background Art

[0002] Power inspection is gradually becoming a hot topic of research today. The safe operation of the power system is related to people's livelihood. A complete power supply system includes a series of related equipment such as power distribution, power generation, power transmission, power consumption, etc. Each link will affect the safe operation of the power system. The most important transmission link at present is studied, mainly studying the behaviors that endanger the safety of the power system on the transmission line (that is, the hidden dangers of abnormalities in the lines, equipment and facilities). The consequences of these hazards to the power system are unimaginable. Therefore, it is a very important link to detect and monitor the safe and effective transmission line and eliminate safety hazards. The power grid company is also constantly conducting in-depth research on it, and constantly improving and innovating the serious impact caused by the transmission line link. At present, the more common defects are broken strand damage, bird's nests, foreign objects and pin shedding. These behaviors that harm the transmission line will affect the safety of power transportation, and more importantly, may threaten the life safety of the staff. At present, the main method of power transmission line inspection in my country is manual inspection. This inspection method can only find defects within the time when the inspector passes by, and cannot truly ensure the accuracy of defect detection. The lack of accurate judgment of defects may lead to false detection, missed detection and wrong classification. These detection methods not only have high requirements on human quality, but also bring a series of resource waste.

[0003] With the advent of the era of artificial intelligence and big data, in response to the above problems, how to use new technologies to replace traditional manual inspections is a very important issue and is in line with the trend of the times. Using image automatic recognition technology to replace traditional manual detection can not only greatly improve the shortcomings of manual recognition (false detection, missed detection), but also improve the accuracy of detection. Moreover, it can greatly improve the real-time performance of detection. Compared with manual inspections, using artificial intelligence technology and image automatic recognition technology to assist power production employees in identifying defects can achieve real-time image defect detection and improve detection accuracy. How to detect defects in real time and efficiently is a hot topic in research. For the research in this invention, it is of great significance to improve the accuracy of identifying transmission line defects through image denoising.

[0004] In the prior art, Chinese patent CN111709891B discloses a training method, an image denoising method, an apparatus and a medium for an image denoising model, which includes constructing a plurality of training pairs, each training pair including a training input image group and a corresponding target image; using the plurality of training pairs to train the image denoising model, but lacks a denoising algorithm optimization method. In particular, due to the use of a manual annotation algorithm, lack of a large number of data sets and comparative experiments, there may be certain deviations from the training result target, resulting in a low accuracy rate.

[0005] Chinese patent CN116645355A discloses a power transmission line monitoring system based on image recognition, including a monitoring module, a positioning module, a communication module, an environmental information acquisition module, an edge computing module and a power supply module. However, since these modular operations themselves may not be applicable to the transmission line scenario, there is a problem of low detection accuracy for the particularity of special power components, which lacks specificity.

[0006] Chinese patent CN117576074A discloses a transmission line insulator defect detection system based on deep learning, including a data acquisition module, a data preprocessing module, a feature extraction module, a model training module, a defect detection module and a result output module. Since the image acquisition source in the transmission line is sensitive, it is very important to expand the data set of the transmission line image. In addition, the lack of clarity of some images has a great impact on the recognition accuracy of the later transmission line.

[0007] Currently, no effective solution has been proposed for the problems in the related technologies. Summary of the invention

[0008] In view of the problems in the related art, the present invention proposes a multi-scale EPLL transmission line image denoising method to overcome the above technical problems existing in the existing related art.

[0009] To this end, the specific technical solution adopted by the present invention is as follows:

[0010] A multi-scale EPLL transmission line image denoising method, the multi-scale EPLL transmission line image denoising method comprising the following steps:

[0011] S1, acquiring a transmission line image, and preprocessing the transmission line image to obtain an image data set;

[0012] S2, extracting image blocks based on the image data set and eliminating DC components;

[0013] S3, calculating the prior distribution parameters of the image block after eliminating the DC component based on the expectation maximization algorithm, and checking the convergence of the likelihood logarithm of the preprocessed power transmission line image;

[0014] S4, combining the semi-quadratic splitting optimization method to minimize the energy function, and alternately optimizing the image blocks and the transmission line images to achieve image denoising;

[0015] Among them, the semi-quadratic splitting optimization method is combined to minimize the energy function, and the image blocks and the transmission line images are optimized alternately to achieve image denoising, including: building a multi-scale framework, converting the noisy image to different color spaces to complete denoising.

[0016] Furthermore, obtaining the transmission line image and preprocessing the transmission line image to obtain the image data set includes the following steps:

[0017] S11, construct a single multi-frame detector, and train the single multi-frame detector using a transmission line image;

[0018] S12, outputting the detection result through the trained single multi-frame detector;

[0019] S13. Process the detection results using non-maximum suppression.

[0020] Furthermore, extracting image blocks based on the image data set and eliminating DC components includes the following steps:

[0021] S21, extracting image blocks of the preprocessed transmission line image;

[0022] S22. Obtain the image block after eliminating the DC component through the expectation maximization algorithm.

[0023] Furthermore, calculating the prior distribution parameters of the image block based on the expectation maximization algorithm and checking the convergence of the likelihood logarithm include the following steps:

[0024] S31, randomly initialize the posterior probability;

[0025] S32, using the initialized posterior probability to calculate the mean, covariance and mixing coefficient of the maximized likelihood function;

[0026] S33, recalculating the posterior probability based on the current parameters;

[0027] S34, calculating the likelihood logarithm of the transmission line image;

[0028] S35, checking the convergence of the likelihood logarithm, if the convergence does not meet the preset standard, returning to S32.

[0029] Furthermore, the semi-quadratic splitting optimization method is combined to minimize the energy function, and the image blocks and the transmission line images are optimized alternately to achieve image denoising, which includes the following steps:

[0030] S41, introducing an auxiliary variable, where the auxiliary variable is each image block in the transmission line image;

[0031] S42, defining a minimum energy function for each image block;

[0032] S43, performing image denoising by alternately optimizing the image blocks and the transmission line image and performing iterations;

[0033] When defining the minimum energy function for each image block, the energy function is defined as:

[0034]

[0035] Where x represents the transmission line image, z i represents the image block, y represents the noise image, y k represents the kth noise value, x k represents the kth pixel;

[0036] Ω represents the image domain, Ω P ∈Ω, ω represents the mixing coefficient;

[0037] β represents a fixed value, P represents a priori value, and σ represents the standard deviation of noise;

[0038] N k Represents the sum of the N DC components of the posterior probability, P i The matrix representing the i-th image patch extracted from the image.

[0039] Furthermore, the image denoising is completed by alternately optimizing the image blocks and the transmission line image and iterating, including the following steps:

[0040] S431, calculating the value of each image block according to the value of the given power transmission line image;

[0041] S432, calculating the value of the power transmission line image according to the value of the image block calculated in S431;

[0042] S433, repeating steps S431-S432, and after each iteration, continuously increasing the fixed value according to the preset fixed value.

[0043] Further, according to the value of the given power transmission line image, calculating the value of each image block includes the following steps:

[0044] S4311, given an image block, eliminating the DC component of the image block, and calculating a mixing weight;

[0045] S4312, selecting a component with the highest weight value;

[0046] S4313, for the selected component value, calculating the maximum a posteriori value through the Wiener filter;

[0047] S4314, adding the deleted DC component to each estimated image block.

[0048] Further, when calculating the value of the power transmission line image according to the value of the image block calculated in S431, the calculation formula of the power transmission line image value is:

[0049]

[0050] Where y represents the noise image, β represents a fixed value, and z represents the image after averaging the overlapping image blocks.

[0051] Furthermore, when implementing image denoising, building a multi-scale framework and applying it to different single-scale denoising algorithms includes the following steps:

[0052] Downsampling of the image is accomplished by extracting the low frequencies from the DCT transform of the image and by computing the inverse DCT only on the low frequencies;

[0053] The DCT transform of the image is used to halve the standard deviation of the Gaussian white noise at each successive scale.

[0054] Furthermore, converting the noisy image to a different color space to perform denoising includes the following steps:

[0055] Denoise each channel of the noisy image;

[0056] Each channel is recombined and the denoising parameters of each channel are adjusted separately by adding a negative feedback regulation mechanism.

[0057] The beneficial effects of the present invention are:

[0058] (1) The present invention can perform denoising based on the prior of image blocks by GMM modeling, and improve the performance of the algorithm by optimizing its parameters. At the same time, a multi-scale framework is added to enable it to be applied to different single-scale denoising algorithms, and it is converted to different color spaces for denoising, thereby reducing artifacts. After experimental comparison and analysis, the performance index of the present invention is higher, the denoising result has less distortion, the detail retention is higher, the clarity of the denoised image in the transmission line is higher, the denoising effect is more obvious, and the edge retention ability is better.

[0059] (2) The present invention uses image recognition to study the defects of power transmission lines. When denoising the power transmission line images, a multi-scale block likelihood logarithmic expectation (EPLL) power transmission line image algorithm based on parameter optimization is used to remove image noise. The block likelihood logarithmic expectation is optimized by a Bayesian image prior algorithm to achieve denoising of the power transmission line image, and a discrete cosine transform (DCT) pyramid is introduced to construct a multi-scale framework, thereby improving the quality of the visual image. In order to reduce the artifact phenomenon generated during image denoising, the image is transformed into different color spaces and denoised channel by channel. The denoising effect is obvious and the accuracy is high.

[0060] (3) The present invention pre-processes the image in advance based on a single multi-frame detector, removes information in the image that is irrelevant to target detection in the pre-processing stage, strengthens the image details, and provides convenience for subsequent processing. When pre-processing the image of the power transmission line, in the early stage of the power transmission line image denoising, the information in the image that is irrelevant to the detection can be deleted, and its information features can be amplified, which is a very important part. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0062] Figure 1 The present invention is a flowchart of a method for denoising a multi-scale EPLL transmission line image according to an embodiment of the present invention. DETAILED DESCRIPTION

[0063] To further illustrate each embodiment, the present invention provides drawings, which are part of the disclosure of the present invention and are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these contents, ordinary technicians in the field should be able to understand other possible implementations and advantages of the present invention. The components in the figures are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0064] According to an embodiment of the present invention, a method for denoising a power line image based on a multi-scale EPLL is provided. The application of the present invention in the power system involves image recognition, image denoising and data algorithm technology, which is used to eliminate information in the image that is irrelevant to target detection, strengthen the details of the image, and amplify its information features, providing convenience for subsequent processing work. The denoising of the power line image is achieved by optimizing the EPLL (optimized block likelihood logarithmic expectation) denoising algorithm (this is a Bayesian-based image prior algorithm). Image priors are a popular tool for image restoration tasks. However, the high dimensionality of the transmission line image makes it very difficult to learn or optimize these priors, and the method of learning image blocks can solve this problem.

[0065] However, applying the prior only to the power line image patches without any control over the entire image is not optimal. In fact, averaging the values ​​obtained for each pixel in the image patch containing it creates new image patches in the constructed image, which can produce errors. The EPLL (Expected Patch Log-Likelihood) method solves this problem. The EPLL algorithm learns the prior (i.e., Gaussian mixture model) from a very large set of image patches taken from several images, rather than using similar image patches of the same image. Therefore, the first step of EPLL is to extract image patches from a dataset of clear images and learn a Gaussian mixture model GMM from them by maximizing the posterior probability. Once the prior is set, for a given noisy image, denoising can be to decompose it into overlapping image patches, denoise each image patch separately, and finally aggregate the results by simple averaging. This likelihood estimation based on overlapping image patches is common in image patch-based algorithms.

[0066] The defect in the modeling based on the transmission line image block is the locality of the model. Even though EPLL strives to work globally on the image, it ignores the interaction and denoising effect in the image when processing small-sized image blocks. The Gaussian mixture model experiment proves that the algorithm can achieve the best performance when the image block size is set to 16x16. Since the processing of each pixel involves a distance of less than 16 pixels, low-frequency noise is not processed. For images with relatively large noise, these residual low-frequency noises become significant, especially in smooth areas. Therefore, the single-scale algorithm runs in a single mode and does not process low-frequency noise, so the denoising effect is not thorough.

[0067] Based on these two observations, a multi-scale framework is constructed by introducing the DCT pyramid in the denoising stage of the power line image, which greatly improves the quality of the visual image. In order to reduce the artifacts generated when denoising the power line image, it is transformed into different color spaces and denoised channel by channel.

[0068] The present invention is further described with reference to the accompanying drawings and specific embodiments. Figure 1As shown, according to the multi-scale EPLL transmission line image denoising method according to an embodiment of the present invention, the multi-scale EPLL transmission line image denoising method comprises the following steps:

[0069] S1, acquiring a transmission line image, and preprocessing the transmission line image to obtain an image data set;

[0070] In order to better detect transmission line defects, images need to be preprocessed in advance. Using a convolutional neural network model to preprocess images can achieve real-time and fast identification without sacrificing accuracy in the detection of transmission line defects. Although the multi-point calculation and analysis speed of YOLO is very good, it sometimes shows poor performance when detecting small defects in transmission lines that are not easily detected by the naked eye. SSD (Single-Shot Detector) is a single-shot structure like YOLO (You Only Look Once, a real-time object detection algorithm), which completes end-to-end model training and extracts a lot of data and features in the middle. SSD can automatically predict and generate ROI areas (Region of Interest) in a single generation step by using a network map similar to RPN (Region Proposal Network, whose main function is to extract candidate boxes) to obtain a series of fixed-size bounding boxes and scores, and finally suppress the results using non-maximum suppression (NMS).

[0071] By using an NMS technique, two highly overlapping bounding box spaces are integrated into one. In the difficult-to-separate sample mining technology, only a subset of negative training samples that fully possess the highest precision training superposition loss can actually be used in each highest training superposition replacement. SSD does not consider the position and boundary classification of each image at the same time, skips the step of region proposal (region proposal, a method of generating candidate regions that may contain target objects), etc., which greatly saves time. Therefore, it is the fastest compared to other models in the transmission line detection link.

[0072] In convolutional neural networks, the convolution layer is the most important part. Its local perception and weight sharing work together with the sub-sampling of the pooling layer to enable it to perform target detection to an excellent state.

[0073] In one embodiment, acquiring a power transmission line image and preprocessing the power transmission line image to obtain an image data set includes the following steps:

[0074] S11. Build a single-shot multi-box detector (SSD) and train the single-shot multi-box detector using transmission line images.

[0075] S12. Output the detection result through the trained single multi-frame detector.

[0076] S13. Use non-maximum suppression to process the detection results, integrate the two highly overlapping bounding box spaces into one, and eliminate information in the image that is irrelevant to target detection.

[0077] S2. Extracting image blocks based on the image data set and eliminating DC components (direct current components).

[0078] In one embodiment, extracting an image block based on an image data set and eliminating a DC component comprises the following steps:

[0079] S21, extracting image blocks of the preprocessed transmission line image;

[0080] S22. Obtain an image block after eliminating the DC component by using an expectation maximization algorithm.

[0081] Assume that the grayscale image (i.e., the overall image of the transmission line) x falls in a grid of size n×m, and describe the image x as a vectorized form, then x k can represent the kth pixel, and P i The matrix representing the i-th image patch extracted from the image, image x i Matrix representation x i =P i x.

[0082] The EPLL algorithm based on image block prior is a Gaussian mixture model. The vector p represents the probability density of the image block and can be expressed as:

[0083]

[0084] Where K represents the number of Gaussian components, f(x|μ k ,∑ k ) is a d-dimensional Gaussian density function, μ k is the mean, Σ k is the covariance, ω k represents the weight, and ∑ k ω k =1.

[0085] The learning process uses the expectation maximization algorithm to calculate N image blocks x that eliminate the DC component i (The DC component corresponds to the average intensity value of the image block), the DC component is the main change direction of the image block, and can be seen from the standard deviation The GMM algorithm mainly focuses on other structures of the image block after eliminating the DC component.

[0086] Given a noise block y i With a local prior p, the maximum a posteriori (MAP) is calculated to estimate the sharp image patch x i , by minimizing the following energy formula:

[0087]

[0088] The goal of the algorithm is to maximize the likelihood function Where N is the DC component, the function has the following parameters (mean, covariance and mixing coefficient of the component). Where σ represents the standard deviation of the noise, and by introducing the posterior probability γ ik To calculate its parameters, the probability is described as:

[0089]

[0090] S3. Calculate the prior distribution parameters of the image block after eliminating the DC component based on the EM algorithm (Expectation Maximization algorithm), and check the convergence of the likelihood logarithm of the preprocessed power line image.

[0091] In one embodiment, calculating the prior distribution parameters of the image block based on the EM algorithm and checking the convergence of the likelihood logarithm includes the following steps:

[0092] S31, randomly initialize the posterior probability γ ik ; where ik represents the kth posterior probability, and j represents the sum of K posterior probabilities.

[0093] S32, M step: Use the initialized posterior probability to calculate the mean of the posterior probability maximizing the likelihood function Covariance and mixing coefficient

[0094]

[0095] in, represents the sum of N DC components of the posterior probability, and T represents the number of optimizations.

[0096] S33, step E: recalculate the posterior probability γ based on the current parameters ik ;

[0097] S34. Calculate the likelihood logarithm of the transmission line image:

[0098]

[0099] Among them, the above formulas μ, Σ, ω represent mean, covariance, and mixing coefficient respectively;

[0100] S35, checking the convergence of the likelihood logarithm, if the convergence does not meet the preset standard, returning to S32.

[0101] It has been experimentally observed that the GMM component μ learned from the dataset of image patches k The average value of is almost 0, so in this invention it is assumed that μ k =0.

[0102] In order to achieve global image processing, it is necessary to calculate the logarithm of the likelihood of the image block (EPLL) over the entire image, while relying on the image model destroyed by noise. For a given image x and image block prior value p, its EPLL value can be defined as:

[0103]

[0104] Where Ω P ∈Ω, is an image block area covered by Ω. Assuming that the image block area in the image is randomly and uniformly distributed, the EPLL value can be interpreted as the logarithm of the likelihood of the image block in the image (directly multiplying the EPLL value of the image by 1 / N), which can be used as the global prior of the image (although the probability of the image distribution in space is not defined). For a given noisy image y, in order to use the image block prior p to minimize the energy function to achieve denoising, the minimized energy function can be defined as:

[0105]

[0106] Where Ω is the image domain, N k Indicates the number of overlapping blocks where pixel k appears (the closer to the image boundary, the larger the N k has a smaller value).

[0107] According to the classification results of the denoised image blocks, the parameters of the Gaussian mixture model are updated. According to the classification results of the image blocks, the parameters of the Gaussian mixture model can be adjusted to better adapt to the characteristics of the image and the noise distribution.

[0108] Optimizing the denoising model: Directly minimizing the energy value (Formula 7) may be very difficult because it depends on the prior value. Therefore, the optimization method of semi-quadratic splitting is used to solve this problem, introducing a set of image blocks As an auxiliary variable, for each image block P in the image i x(P i The probability of occurrence of each image block), when defining the minimum energy function E generated by each image block, the energy function is defined as:

[0109]

[0110] Where x represents the transmission line image, z i represents the image block, y represents the noise image, y k represents the kth noise value, x k represents the kth pixel; Ω represents the image domain, Ω P ∈Ω, ω represents the mixing coefficient; β represents a fixed value, P represents a priori value, σ represents the noise standard deviation; N k Represents the sum of N DC components of the posterior probability, P i The matrix representing the i-th image patch extracted from the image.

[0111] When β→+∞, the image block P can be set i x is equal to the auxiliary variable {z i}, and the solution of formula (8) can be considered as an approximate solution of formula (7). For a fixed value of β, we can continuously iterate to optimize formula (8), first solving for x while keeping {z i} constant, then keep the newly calculated value of x constant and calculate {z i The algorithm uses the noise image y to initialize the value of x, which takes about six iterations to solve. After each iteration, the value of β is continuously increased according to the pre-set value of β.

[0112] S4. Combine the semi-quadratic splitting optimization method to minimize the energy function, and alternately optimize the image blocks and the transmission line images to achieve image denoising.

[0113] From formula (5), we can see that minimizing {z i The process of {z} is separable, which means that the image patch {z i} and is independent of other image blocks. Each process can be regarded as a process with a prior value p and a noise value Under the condition of i}MAP.

[0114] In one embodiment, the image denoising is implemented by minimizing the energy function by combining the semi-quadratic splitting optimization method and alternately optimizing the image blocks and the transmission line image, and the following steps are included:

[0115] S41, introducing an auxiliary variable, where the auxiliary variable is each image block in the transmission line image;

[0116] S42, defining a minimum energy function for each image block;

[0117] S43, image denoising is completed by alternately optimizing the image blocks and the transmission line image and iterating.

[0118] In one embodiment, the image denoising is performed by alternately optimizing the image blocks and the transmission line image and iteratively, including the following steps:

[0119] S431, calculating the value of each image block according to the value of the given power transmission line image;

[0120] S432, calculating the value of the power transmission line image according to the value of the image block calculated in S431;

[0121] S433, repeating steps S431-S432, and after each iteration, continuously increasing the fixed value according to the preset fixed value.

[0122] In one embodiment, according to the value of a given power transmission line image, calculating the value of each image block comprises the following steps:

[0123] S4311: Given an image block P i x, eliminate the DC component of the image block and calculate the mixing weight γ ik =P(k|P i x);

[0124] S4312, select the component with the highest weight value

[0125] S4313, for the selected component value The MAP value is calculated by the Wiener filter, where i represents the Wiener filter weight;

[0126]

[0127] S4314, adding the deleted DC component to each estimated image block.

[0128] For {z i} value, the value of x can be solved by the following formula:

[0129]

[0130] P i T z i represents the adjoint matrix of P, creating a new image by replacing the image block z i Placed in the appropriate position, the rest of the image is set to zero, and the matrix is a diagonal matrix of size nm×nm. Each element in the diagonal matrix is ​​associated with a pixel and is equal to the number of image blocks. z is set to the overlapping image block z. i The image after averaging can be defined as follows:

[0131]

[0132] In one embodiment, when calculating the value of the power transmission line image according to the value of the image block calculated in S431, the calculation formula of the power transmission line image value is:

[0133]

[0134] That is, formula (12) is a redefinition of formula (10), where y represents the noise image, β represents a fixed value, and z represents the image after averaging the overlapping image blocks. This means that the optimization of g in each step is simply the aggregation of image blocks, resulting in a weighted average between the noise image y and the image z.

[0135] Multi-scale framework construction:

[0136] When implementing image denoising, in order to improve the performance of the algorithm, the present invention constructs a multi-scale framework and applies it to different single-scale EPLL denoising algorithms;

[0137] This has two limitations. First, in order to be able to effectively adapt each original single-scale algorithm, each layer needs to be represented at multiple scales to maintain Gaussian white noise. Second, the output image needs to be efficiently reconstructed from the denoising results at different scales. A solution to these two limitations is to use a modified DCT pyramid. DCT stands for discrete cosine transform. Downsampling of the image is simply done by extracting the low frequencies from the DCT transform of the image and by calculating the inverse DCT (IDCT) only on the low frequencies, so that each layer of the pyramid has half the width and half the height of the previous layer. In addition, the DCT transform of the image is such that the standard deviation of the Gaussian white noise is halved at each successive scale. Therefore, there is no need to make specific adjustments to the initial single-scale denoising algorithm to denoise the coarse layer. The low frequencies of the layer can be replaced by the frequencies of the coarser layer.

[0138] The improved DCT transform of the two-dimensional image f(x,y) can be expressed as:

[0139]

[0140] Where a represents the probability of the original signal point, and its inverse transform (IDCT) can be expressed as:

[0141]

[0142] Wherein, F(x, y) represents the coefficient after DCT transformation, M, N represent the number of points of the original signal, u=0,1...M-1, v=0,1...N-1.

[0143] The results without and with the multi-scale framework are shown in Table 1. The size of these images is 768×576, and the multi-scale method uses 3 scales.

[0144] Table 1: Impact of multi-scale framework

[0145]

[0146] For small amounts of noise, denoising is easy and there is no need to add complex multi-scale methods because ringing artifacts will deteriorate the denoising results. However, once σ = 30, these artifacts due to the influence of noise are negligible.

[0147] In one embodiment, when implementing image denoising, constructing a multi-scale framework and applying it to different single-scale denoising algorithms includes the following steps:

[0148] Downsampling of the image is accomplished by extracting the low frequencies from the DCT transform of the image and by computing the inverse DCT only on the low frequencies;

[0149] The DCT transform of the image is used to halve the standard deviation of the Gaussian white noise at each successive scale.

[0150] Image Denoising:

[0151] De-noising is done by converting the noisy image into a different color space.

[0152] The present invention denoises each channel R, G and B of the noise image independently and recombines them together. By adding a negative feedback regulation mechanism, the denoising parameters of the channels R, G and B are adjusted respectively, so as to obtain a better denoising effect.

[0153] Using R, G and B channels to denoise individually will produce noticeable color artifacts. Usually noisy images in RGB color space can be converted to images in different color spaces, and then denoising each channel independently will reduce these artifacts.

[0154] The present invention tests two conversions: YUV color space conversion and OPP color space conversion.

[0155]

[0156] Therefore, setting σ is still the standard deviation of the noise on each channel. The comparative denoising performance of these different color transforms is shown in Table 2.

[0157] Table 2: Effects of different color channels

[0158]

[0159]

[0160] From Table 2, we can see that the PSNR (Peak Signal-to-Noise Ratio) is higher in YUV and OPP color spaces than in RGB color space. The first channel of the OPP transform is the average of the three colors, which captures the geometry of the image patch. The average value improves the signal-to-noise ratio, thus achieving better results in this color space. The remaining channels represent the color content of the image patch. And A opp is an orthogonal matrix that is uncorrelated with the noise and A YUV No, so the algorithm performs better in the OPP color space.

[0161] In one embodiment, converting a noisy image to a different color space to perform denoising includes the following steps:

[0162] Denoise each channel R, G and B of the noisy image;

[0163] Each channel is regrouped together and the denoising parameters of each channel are adjusted separately by adding a negative feedback regulation mechanism.

[0164] Prediction results:

[0165] The present invention uses an image database consisting of JPG images of size 768×576. In order to treat them as noise-free images, the present invention reduces these images by a factor of 2. For this purpose, the images are filtered with a Gaussian kernel with a standard deviation of 1:4, selecting one pixel out of 4. Among the 419 images of size 384×288, about 70% of the images are uniformly selected for learning, and the remaining 30% become the test dataset.

[0166] The image patch is represented by a vector From this database, approximately 30×10 6 image blocks and removed their DC components, and then randomly selected 2×10 6 Patch to implement the EM algorithm.

[0167] The noise values ​​are set to σ = 20, σ = 30 and σ = 50 respectively, and the size of the image block dataset is N = 2 10 6 , the number of steps s = 1. The experimental results are respectively EPLL algorithm (EPLL), parameter optimized EPLL algorithm (Re-EPLL), and optimized EPLL algorithm of the present invention (MC-EPLL). Both Re-EPLL algorithm and MC-EPLL algorithm are set according to the parameters optimized by the present invention.

[0168] Table 3: Tests Figure 1 A PSNR value of -3

[0169]

[0170] It can be seen from Table 3 that the EPLL algorithm after parameter optimization has a more thorough denoising and less image distortion. The EPLL algorithm with a multi-scale framework and converted to the OPP color space has less artifacts and a better image denoising effect.

[0171] Algorithm comparison and analysis:

[0172] In order to better evaluate the performance of the present invention, it is compared with the most advanced methods currently available: BM3D (block matching and three-dimensional filtering algorithm), Non-local Bayes algorithm (non-local Bayes algorithm), DCT algorithm (discrete cosine transform algorithm), K-SVT algorithm (K singular value threshold algorithm) and wavelet denoising algorithm.

[0173] Some performance indicators are used for evaluation. The performance evaluation indicators of denoising effect are mainly verified by peak signal to noise ratio (PSNR), structural similarity (SSIM) and edge preservation index (EPI). Peak signal to noise ratio (PSNR) is the ratio between the maximum possible value of the signal and the distortion noise power that affects the image quality. It is based on the error between corresponding pixels, that is, based on error-sensitive image quality evaluation. It is the most common and widely used objective image evaluation indicator. The unit of PSNR is Db. The larger the value, the smaller the distortion. It can be expressed by the following formula:

[0174]

[0175] Among them, f and It represents the original image and the restored image, MAX is the maximum value of the color of the image point, and MSE (mean square error) represents the mean of the squared error between the damaged image and the reconstructed image. Therefore, a lower mean square error value can ensure a higher PSNR value.

[0176] Although PSNR is now more commonly used because it is mathematically convenient and straightforward to calculate, it does not match the visual observation of image quality. An important limitation is that it strictly relies on numerical analysis and does not actually consider the level of the human visual system, while structural similarity (SSIM) is a good performance indicator for visually comparing images.

[0177]

[0178] Among them, only f, Respectively represent the two images to be compared. hour, This means that the image reconstruction quality is better. f and is the average intensity of the original image and the restored image, σ f and is the standard deviation of the original image and the restored image, Represents the covariance between the original image and the restored image. c1 and c2 are very small constants that can avoid the denominator being zero. The higher the structural similarity, the richer the structural information of the image and the higher the quality of image reconstruction.

[0179] The edge preservation coefficient indicates the ability of the filter to preserve the edge details of the image after processing. The higher the edge preservation coefficient value, the higher the preservation ability, and the better the processing performance of this filter. The calculation formula is:

[0180]

[0181] Among them, m is the number of image pixels, G R1 and G R2 The closer EPI is to 1, the stronger the edge preservation ability of the algorithm is.

[0182] The denoising effect of the present invention in the power transmission line is more thorough, among which five algorithms, BM3D, Non-local Bayes, DCT, K-SVT and Our multi-scale EPLL denoising, are compared in experiments under the same noise condition.

[0183] The improved algorithm can completely remove some noise, and no obvious loss will occur in some of its feature maps. In defective images, there are fewer artifacts and the image details are more preserved. Therefore, the denoising effect of the multi-scale EPLL transmission line image based on parameter optimization is more obvious. Then the comparative experiment of each algorithm index is shown in Table 4 below.

[0184] Table 4: Denoising index value comparison

[0185]

[0186]

[0187] Table 3-4 reflects the average indicators of the above-mentioned transmission line images. The denoised images of the transmission lines of the present invention have higher clarity and better edge retention ability.

[0188] Glossary:

[0189] EPLL, optimized block-likelihood log-expected.

[0190] GMM, Gaussian mixture model.

[0191] DCT, discrete cosine transform.

[0192] PSNR, peak signal to noise ratio.

[0193] SSIM, structural similarity.

[0194] EPI, edge preservation index.

[0195] In summary, the present invention can perform denoising based on the prior of the image block by GMM modeling, and improve the performance of the algorithm by optimizing its parameters. At the same time, a multi-scale framework is added to apply it to different single-scale denoising algorithms, and it is converted to different color spaces for denoising, reducing the artifact phenomenon. After experimental comparison and analysis, the performance index of the present invention is high, the distortion of the denoising result is small, the detail retention is high, the clarity of the denoised image in the transmission line is high, the denoising effect is more obvious, and the edge retention ability is good. The present invention adopts image recognition to study the defects of the transmission line. When denoising the transmission line image, the multi-scale block likelihood logarithm expectation EPLL transmission line image algorithm based on parameter optimization is used to remove image noise. The block likelihood logarithm expectation is optimized by the Bayesian image prior algorithm to achieve denoising of the transmission line image, and the discrete cosine transform DCT pyramid is introduced to construct a multi-scale framework, which improves the quality of the visual image. In order to reduce the artifact phenomenon generated during image denoising, by transforming it into different color spaces, denoising is performed channel by channel, and the denoising effect is obvious and the accuracy is high. The present invention pre-processes the image in advance based on a single multi-frame detector, removes information in the image that is irrelevant to target detection in the pre-processing stage, strengthens the details of the image, and provides convenience for subsequent processing. When pre-processing the image of the power transmission line, in the early stage of the power transmission line image denoising, the information in the image that is irrelevant to the detection can be deleted, and its information features can be amplified, which is a very important part.

[0196] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A multi-scale EPLL transmission line image denoising method, characterized in that: The multi-scale EPLL transmission line image denoising method comprises the following steps: S1, acquiring a transmission line image, and preprocessing the transmission line image to obtain an image data set; S2, extracting image blocks based on the image data set and eliminating DC components; S3, calculating the prior distribution parameters of the image block after eliminating the DC component based on the expectation maximization algorithm, and checking the convergence of the likelihood logarithm of the preprocessed power transmission line image; S4, combining the semi-quadratic splitting optimization method to minimize the energy function, and alternately optimizing the image blocks and the transmission line images to achieve image denoising; Among them, the semi-quadratic splitting optimization method is combined to minimize the energy function, and the image blocks and the transmission line images are optimized alternately to achieve image denoising, including: building a multi-scale framework, converting the noisy image to different color spaces to complete denoising.

2. The method for denoising a multi-scale EPLL transmission line image according to claim 1, characterized in that: The step of acquiring the transmission line image and preprocessing the transmission line image to obtain the image data set comprises the following steps: S11, construct a single multi-frame detector, and train the single multi-frame detector using a transmission line image; S12, outputting the detection result through the trained single multi-frame detector; S13. Process the detection results using non-maximum suppression.

3. The method for denoising a multi-scale EPLL transmission line image according to claim 2, characterized in that: The method of extracting image blocks based on the image data set and eliminating DC components comprises the following steps: S21, extracting image blocks of the preprocessed transmission line image; S22. Obtain the image block after eliminating the DC component through the expectation maximization algorithm.

4. The method for denoising a multi-scale EPLL transmission line image according to claim 3, characterized in that: The method of calculating the prior distribution parameters of the image block based on the expectation maximization algorithm and checking the convergence of the likelihood logarithm includes the following steps: S31, randomly initialize the posterior probability; S32, using the initialized posterior probability to calculate the mean, covariance and mixing coefficient of the maximized likelihood function; S33, recalculating the posterior probability based on the current parameters; S34, calculating the likelihood logarithm of the transmission line image; S35, checking the convergence of the likelihood logarithm, if the convergence does not meet the preset standard, returning to S32.

5. The method for denoising a multi-scale EPLL transmission line image according to claim 4, characterized in that: The method of minimizing the energy function by combining the semi-quadratic splitting optimization method and alternately optimizing the image blocks and the transmission line images to achieve image denoising includes the following steps: S41, introducing an auxiliary variable, where the auxiliary variable is each image block in the transmission line image; S42, defining a minimum energy function for each image block; S43, performing image denoising by alternately optimizing the image blocks and the transmission line image and performing iterations; When defining the minimum energy function for each image block, the energy function is defined as: Where x represents the transmission line image, z i represents the image block, y represents the noise image, y k represents the kth noise value, x k represents the kth pixel; Ω represents the image domain, Ω P ∈Ω, ω represents the mixing coefficient; β represents a fixed value, P represents a priori value, and σ represents the standard deviation of noise; N k Represents the sum of the N DC components of the posterior probability, P i The matrix representing the i-th image patch extracted from the image.

6. The method for denoising a multi-scale EPLL transmission line image according to claim 5, characterized in that: The method of alternately optimizing the image blocks and the transmission line image and completing the image denoising after iteration includes the following steps: S431, calculating the value of each image block according to the value of the given power transmission line image; S432, calculating the value of the power transmission line image according to the value of the image block calculated in S431; S433, repeating steps S431-S432, and after each iteration, continuously increasing the fixed value according to the preset fixed value.

7. The method for denoising a multi-scale EPLL transmission line image according to claim 6, characterized in that: The step of calculating the value of each image block according to the value of the given power transmission line image comprises the following steps: S4311, given an image block, eliminating the DC component of the image block, and calculating a mixing weight; S4312, selecting a component with the highest weight value; S4313, for the selected component value, calculating the maximum a posteriori value through the Wiener filter; S4314, adding the deleted DC component to each estimated image block.

8. The method for denoising a multi-scale EPLL transmission line image according to claim 6, characterized in that: When calculating the value of the power transmission line image according to the value of the image block calculated in S431, the calculation formula of the power transmission line image value is: Where y represents the noise image, β represents a fixed value, and z represents the image after averaging the overlapping image blocks.

9. The method for denoising a multi-scale EPLL transmission line image according to claim 8, characterized in that: The construction of the multi-scale framework comprises the following steps: Downsampling of the image is accomplished by extracting the low frequencies from the DCT transform of the image and by computing the inverse DCT only on the low frequencies; The DCT transform of the image is used to halve the standard deviation of the Gaussian white noise at each successive scale.

10. The method for denoising a multi-scale EPLL transmission line image according to claim 9, characterized in that: The process of converting the noisy image into a different color space to perform denoising comprises the following steps: De-noise each channel of the noisy image; Each channel is recombined and the denoising parameters of each channel are adjusted separately by adding a negative feedback regulation mechanism.