Image denoising method for automatic optical detection system in electronic industry

Abstract

The invention provides an image denoising method for an automatic optical detection system in the electronic industry, and aims to efficiently remove image noise and reserve details. The method comprises the following steps of: firstly, constructing acquired images into a plurality of similar group tensors through blocking and similarity grouping; secondly, modeling and processing spatial and spectral characteristics of similar group tensors by utilizing low-rank tensor decomposition; secondly, exploring the global low-rank property of the image through a three-direction logarithmic truncation tensor nuclear norm (3TLogTNN) so as to suppress noise and keep the global structure consistency; and finally, constructing an optimization objective function, and realizing image denoising and detail reservation in combination with a non-local two-factor regularizer and a global regularizer. According to the method, non-local similarity and global low-rank characteristics are effectively combined, the image denoising effect and algorithm efficiency are improved, and the method is suitable for multispectral images and high-precision detection scenes and can be widely applied to an AOI system in the electronic industry.

Description

Technical Field

[0001] The present invention relates to the field of image processing and industrial intelligent manufacturing, and in particular to an image denoising method used in an automatic optical inspection system in the electronics industry. The method utilizes global and non-local low-rank tensor decomposition technology to significantly improve the detection accuracy and efficiency of the AOI system by effectively suppressing noise and retaining image details. Background Art

[0002] Automatic Optical Inspection (AOI) technology is a non-contact inspection method widely used in the electronics manufacturing industry. It is mainly used to detect defects in tiny components such as circuit boards and semiconductor chips. In the electronics manufacturing process, due to the small size and complex structure of components, traditional manual inspection methods have been difficult to meet the requirements of high precision and high efficiency. The AOI system uses computer vision technology to collect images of the object to be tested using a camera, and then automatically analyzes the image through image processing and pattern recognition algorithms to identify whether there are defects. This technology can significantly improve production efficiency and quality, and is an indispensable quality control method in the current electronics manufacturing industry.

[0003] In the AOI system, image processing is one of its core modules. This module usually includes image acquisition, image denoising, feature extraction, image matching and other links. The quality of image acquisition and denoising directly affects the accuracy of the subsequent inspection process. Since the AOI system has high requirements for image quality, it is often interfered by external factors such as camera performance, changes in light conditions, electronic noise, etc. during the image acquisition process, resulting in different degrees of noise in the image. The presence of noise will reduce the clarity of the image, blur details, and even cover up defect features, seriously affecting the detection algorithm's accurate identification of defects. Therefore, image denoising technology has become one of the key steps in the image processing link of the AOI system.

[0004] Currently, common image denoising methods can be divided into two categories: global denoising and local denoising.

[0005] 1. Global denoising method

[0006] Global denoising methods usually analyze the overall structural characteristics of the image, such as the sparsity or low-rank characteristics of the image, and use the redundant information in the image to restore the noise-contaminated image. Global denoising methods are effective in removing large-area, uniformly distributed noise. Common global denoising methods include Fourier transform, discrete wavelet transform, low-rank matrix decomposition, etc. These methods suppress noise by processing the overall structure of the image in the frequency domain or spatial domain. However, global denoising methods often have weak retention capabilities for image details and are prone to blurred edges. Especially when processing images with a large amount of details, some important information may be lost, thus affecting the accuracy of detection.

[0007] 2. Local denoising method

[0008] Local denoising methods are mainly based on neighborhood filtering algorithms, which reduce noise by calculating the average or weighted value of pixels in the neighboring area of ​​each pixel. Local denoising methods include mean filtering, median filtering, Gaussian filtering, etc. These methods perform well in protecting image edge details and can effectively remove high-frequency noise in local areas. However, local denoising methods are limited to processing noise within the pixel neighborhood, and are less effective in removing large-area, low-frequency noise. In addition, local denoising methods have high computational complexity, and take a long time to operate when processing high-resolution images or large batches of images.

[0009] In the application of AOI system in the electronics manufacturing industry, although the traditional image denoising method has improved the image quality to a certain extent, it has some obvious shortcomings, such as:

[0010] Non-local similarity is common in natural images, that is, similar blocks at different locations in the image are highly correlated. For example, structures such as textures and edges in a certain area of ​​the image usually appear repeatedly in other areas. Non-local similarity can be used to enhance the effect of image denoising because it can integrate similar information at different locations to further suppress noise. However, most traditional denoising methods are based on the similarity of local neighborhoods and only process within a limited pixel range, failing to fully utilize the non-local information of the image, which results in limited image denoising effects, especially in detail areas such as edges and textures.

[0011] Most of the image denoising methods commonly used in AOI systems treat images as two-dimensional matrices for processing. For multi-channel RGB images or multispectral images, each channel is usually processed separately, ignoring the correlation between channels. Although this processing method can simply and directly reduce noise, it fails to fully utilize the complementary information between different channels of the image. Especially in multispectral images, the data correlation between channels is strong, and traditional methods cannot effectively utilize these multi-channel connections, resulting in the denoising effect cannot meet the high-precision requirements.

[0012] In industrial production, AOI systems need to process a large number of images collected in real time, so the denoising method must not only have a good denoising effect, but also meet the real-time requirements. However, many advanced denoising algorithms improve image quality while increasing computational complexity. These methods usually require large computing resources and long processing time, which is difficult to meet the AOI system's demand for fast processing.

[0013] In AOI applications, detail preservation is crucial because tiny defects often only exist in certain subtle structures of the image. Excessive denoising may cause loss of image details, thereby masking product defects. Global denoising methods and local denoising methods have certain limitations in processing image edges and details. For example, global denoising is prone to cause edge blurring, while local denoising may not be able to fully restore image details. Existing image denoising methods still need to be improved in terms of the balance between detail preservation and noise suppression.

[0014] With the increasing application of multi-dimensional data such as multispectral and hyperspectral images in AOI systems, how to retain the multi-dimensional information of images while denoising has become a new challenge. Tensors are a tool for representing multi-dimensional data, which can simultaneously express image information in multiple dimensions such as space, channels, and spectra. Compared with traditional matrix representation, tensor models have natural advantages in processing multi-dimensional image data. However, most current denoising methods are still based on two-dimensional matrices for processing, and fail to effectively utilize the multi-dimensional representation characteristics of tensors, resulting in the multi-dimensional structural information of the image not being fully retained during the denoising process, thereby affecting the detection accuracy of the AOI system.

[0015] In summary, the existing AOI image denoising methods have obvious deficiencies in non-local similarity utilization, channel correlation mining, real-time performance, detail retention, and multi-dimensional information processing. Faced with the increasing quality inspection requirements of the electronics manufacturing industry, an improved denoising method is urgently needed that can effectively overcome the above-mentioned defects in the existing technology, improve the image denoising effect of the AOI system in practical applications, and provide clearer and more accurate image data for high-precision defect detection. Summary of the invention

[0016] In order to solve the above problems in the prior art, the present invention proposes an image denoising method for an automatic optical inspection system in the electronics industry, which is characterized by comprising the following steps:

[0017] 1) Divide the collected images into blocks and group them according to similarity, including the following sub-steps:

[0018] Dividing the acquired image into a number of sub-regions to form a plurality of three-dimensional image blocks;

[0019] Calculate the similarity between each three-dimensional image block and other image blocks by using a similarity measurement method;

[0020] According to the similarity, similar image blocks are grouped to form several similar groups containing multiple image blocks;

[0021] Construct a three-dimensional tensor for each similar group of image patches;

[0022] 2) Low-rank tensor modeling of similar blocks, including the following sub-steps:

[0023] The similarity group tensor G i Decomposed into spatial factor tensor B i and the spectral matrix A i , where i is the similarity group index currently being processed, indicating the i-th similarity group;

[0024] For the spatial factor tensor B i Processing to extract spatial features of the image;

[0025] For the spectral matrix A i Processing is performed to smooth spectral features;

[0026] Synthesize the processed spatial factor tensor B i and the spectral matrix A i , generate the denoised similar tensor G i ;

[0027] 3) Perform global low-rank exploration on the denoised image, including the following sub-steps:

[0028] Compute multimodal low-rank properties of images;

[0029] Use global regularization methods to explore the global low-rank nature of images;

[0030] 4) Construct an image denoising model to achieve noise suppression and detail retention by optimizing the objective function, where the optimization objective function is:

[0031]

[0032] in:

[0033] Y: collected original image data;

[0034] The denoised image represents the noise-free image restored by the denoising model;

[0035] s: sparse noise term, used to compensate for the high-frequency noise remaining in the original image;

[0036] G i : The i-th similarity group tensor obtained by tensor decomposition;

[0037] B i : The spatial factor tensor of the i-th similarity group, used to extract spatial features;

[0038] A i : The spectral matrix of the i-th similarity group is used to extract spectral features;

[0039] Spectral matrix A i A smoothed version of , used to remove noise in the spectral dimension;

[0040] D3A i : Spectral matrix A i The differential representation of is used to further suppress spectral noise;

[0041] ρ: regularization parameter, used to control the weight of the three-way logarithmic truncated tensor nuclear norm (3TLogTNN) regularization term;

[0042] μ: spectral smoothing regularization parameter, used to control the weight of the spectral difference regularization term;

[0043] γ: sparse regularization parameter, used to control the regularization weight of the sparse noise term S;

[0044] The square of the Frobenius norm, used to measure the reconstruction error and regularization term;

[0045] ‖·‖ 3DTLogTNN : Three-directional logarithmic truncated tensor nuclear norm, used to explore the global low rank of the image;

[0046] ‖·‖1: norm, used for sparse regularization;

[0047] n: The total number of similar groups, used to represent the number of all similar blocks that constitute the complete image.

[0048] The non-local two-factor regularizer includes the spatial difference image D1B i and D2B i constraints.

[0049] The spectral smoothing includes the following steps: i Perform singular value decomposition.

[0050] The calculation of the three-way logarithmic truncated tensor nuclear norm (3TLogTNN) includes the TLogMNN calculation of the mode-k slice of the tensor, and its formula is as follows:

[0051]

[0052] in:

[0053] X: input tensor slice;

[0054] σ i : The i-th singular value of the input tensor slice X, arranged in descending order;

[0055] ∈: a small constant used to avoid taking logarithms of zero values, usually a very small positive number;

[0056] m: The number of singular values ​​retained, which means that the logarithm operation is performed on the first m singular values.

[0057] The objective function of the image denoising model includes the following constraints:

[0058]

[0059] in:

[0060] G i : represents the low-rank tensor representation of the i-th similar group obtained by tensor decomposition;

[0061] B i : The spatial factor tensor of the i-th similarity block, used to extract the spatial features of the image;

[0062] A i : The spectral matrix of the i-th similar block is used to extract the spectral features of the image;

[0063] ×3: represents tensor multiplication along the third mode (channel direction);

[0064] The complete image after denoising is represented by a low-rank tensor G of all n similar blocks i Reconstruction obtained;

[0065] n: The total number of similar blocks, used to indicate the total number of similar groups formed by image blocks.

[0066] Beneficial effects:

[0067] By combining non-local similarity and global low-rank characteristics, an image denoising model based on low-rank tensor decomposition and regularization method is proposed, which effectively suppresses image noise while retaining the spatial details and spectral information of the image. The block segmentation and similarity group construction technology are used to fully utilize the non-local similarity of the image to improve the accuracy of denoising; the three-directional logarithmic truncated tensor nuclear norm (3TLogTNN) is used to explore the global low-rank of the image, improve the denoising effect and reduce the loss of details; the optimization objective function is combined with a variety of regularizers to further enhance the robustness and efficiency of the model, which can significantly improve the quality and reliability of image processing in the AOI system. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present application, but do not constitute an improper limitation of the present invention. In the drawings:

[0069] Figure 1 : shows the overall flow chart of the image denoising method. DETAILED DESCRIPTION

[0070] The present invention will be described in detail below in conjunction with the accompanying drawings and specific embodiments, wherein the illustrative embodiments and descriptions are only used to explain the present invention but are not intended to limit the present invention.

[0071] This embodiment provides an image denoising method suitable for the automatic optical inspection system in the electronics industry, which achieves denoising effect and retains image details by performing block segmentation, grouping, low-rank tensor modeling and global low-rank exploration on the collected images. The specific implementation steps are as follows:

[0072] Step 1: Image acquisition, segmentation and grouping

[0073] 1.1 Image acquisition

[0074] In this embodiment, the camera in the AOI system is first used to capture the image of the target to be measured, and an RGB image or a multispectral image of the target is obtained. The collected image data is recorded as Y, which contains multiple channels. For RGB images, three channels (red, green and blue) are usually included, while multispectral images may contain more spectral channels. The data size of each channel of the image is M×N, that is, the number of pixels of the image in the horizontal and vertical directions. Since the AOI system has high requirements for image clarity and details, ambient light interference should be reduced as much as possible during image acquisition, and the camera equipment should be ensured to have sufficient resolution to ensure the accuracy and effectiveness of subsequent image denoising processing.

[0075] 1.2 Image segmentation

[0076] In order to better denoise the image, in this embodiment, the collected RGB image or multispectral image is divided into blocks according to a preset patch size. Specifically, each image is divided into n three-dimensional image blocks, and the size of each image block is Where s represents the spatial size of the image block (i.e., the number of pixels contained in each image block), and k represents the number of channels of the image. Through this block division method, the entire image can be divided into several small areas, so that each image block contains richer local information. This local information can not only improve the resolution of image details, but also provide a basis for subsequent similarity analysis and denoising.

[0077] The advantage of block processing is that it can divide a large image into small blocks that are easy to process, making it easier for the algorithm to perform a detailed analysis of the local features of the image. Each block can be regarded as a small independent unit, which is convenient for block-by-block similarity calculation and tensor modeling in subsequent steps. At the same time, this block method also makes the denoising process more flexible, and can optimize the noise characteristics of each block, thereby improving the overall denoising effect.

[0078] 1.3 Similarity calculation and grouping

[0079] After the block processing, the similarity between each image block and other image blocks is calculated to group similar image blocks. In this embodiment, the similarity is measured using the Euclidean distance as the standard. Specifically, an image block G is selected. i As the benchmark block currently being processed, the Euclidean distance between it and all other image blocks is calculated. Euclidean distance is a commonly used similarity measurement method that can measure the similarity between two image blocks. The smaller the distance, the higher the similarity of the image blocks.

[0080] By i By sorting the Euclidean distance of the image blocks, other image blocks that are most similar to the current image block can be identified. The first b-1 most similar image blocks in the sorting result are selected and compared with the reference image block G i The similarity group is composed of b image blocks, each of which contains b image blocks, where i represents the similarity group index currently being processed. The purpose of similarity grouping is to aggregate image blocks with similar structure or texture features in the image, so as to facilitate the subsequent denoising process through the statistical characteristics of similar blocks.

[0081] This similarity grouping method can make full use of the non-local self-similarity characteristics of the image, so that the information of multiple similar image blocks can be considered simultaneously during the denoising process, improving the denoising effect and the retention of image details. At the same time, this denoising method based on similarity groups can effectively reduce the noise difference of image blocks and improve the stability of processing.

[0082] 1.4 Tensor Construction

[0083] After the similarity grouping is completed, the image blocks of each similarity group are expanded along the third mode (i.e., the channel direction) to transform the three-dimensional structure of each image block into a two-dimensional matrix structure. Specifically, for the b image blocks in each similarity group, each image block is first expanded along the channel direction to obtain b matrices of size s×k. Then, these b matrices are stacked along the third mode to form a new tensor G of size k×s×b i , where i represents the index of the current similarity group.

[0084] Through this tensor construction process, each similarity group is transformed into a three-dimensional tensor Gi , where each dimension represents channel, space, and similarity respectively. This tensor structure can capture the characteristic information of image patches in multiple dimensions, so that the subsequent low-rank tensor modeling can simultaneously utilize the multi-channel information, spatial structure, and similarity characteristics of the image.

[0085] The advantage of tensor construction is that it can preserve the multi-dimensional features of the image and provide a rich source of information for low-rank modeling. Especially in the denoising of multispectral images, the tensor form can effectively utilize the correlation between different channels to improve the denoising effect. In addition, the structure after tensor construction is more suitable for low-rank tensor decomposition algorithms, making the denoising process more accurate and robust, which helps to retain image details and effectively suppress noise.

[0086] In summary, this embodiment lays the foundation for subsequent image denoising through steps such as image acquisition, block division, similarity calculation and grouping, and tensor construction. By utilizing non-local self-similarity and the multi-dimensional features of tensors, this method can maintain the integrity of image details during the denoising process, thereby improving the detection accuracy of the AOI system.

[0087] Step 2: Low-rank tensor modeling of similar blocks

[0088] 2.1 Modular Three-Tensor Matrix Decomposition

[0089] In this step, in order to make full use of the low-rank characteristics of similar blocks, the non-local similarity tensor G i Perform modular three-dimensional tensor matrix decomposition. Specifically, the tensor G i Decomposed into two parts: spatial factor tensor B i and the spectral matrix A i , thereby extracting the spatial information and spectral information of the image. The decomposition process can be expressed as:

[0090] G i ≈B i ×3A i

[0091] The symbol ×3 represents the tensor multiplication operation along the third mode (i.e., channel direction). Through this decomposition process, the spatial structure and spectral information of the image can be effectively separated, providing a basis for subsequent denoising. Spatial factor tensor B i Represents the spatial structural characteristics of the image block, and the spectral matrix A i This decomposition allows the spatial and spectral information of the image to be processed separately during the denoising process, thereby improving the denoising effect and the retention of image details.

[0092] 2.2 Spatial Difference

[0093] In order to further enhance the spatial structural features of the image, in this step, the spatial factor tensor B i Specifically, B i Differential processing is performed along the first mode (horizontal direction) and the second mode (vertical direction) to obtain a spatial differential image D1B i and D2B i Through the difference operation, the edge information of the image block can be effectively captured. Edge information is usually an important detail feature in the image, so retaining this information helps to enhance the quality of the denoised image and make the denoised image visually clearer.

[0094] Spatial difference image D1B i and D2B i It can highlight the edges and structural change areas in the image, which helps to enhance the spatial consistency of the image in the subsequent denoising process. Through this difference operation, not only can the noise be effectively removed, but also the spatial details of the image can be preserved, making the denoised image closer to the original image.

[0095] 2.3 Spectral smoothing

[0096] In this step, the spectral matrix A i Singular value decomposition (SVD) is performed and spectral smoothing is performed. Specifically, the spectral matrix A is transformed into i Decomposed into a combination of singular values ​​and singular vectors. In order to achieve spectral smoothing effect, the singular values ​​are screened and only the main singular values ​​are retained, and the smaller singular values ​​are removed. The retained singular values ​​can effectively represent the main spectral information of the image, while the smaller singular values ​​usually contain noise components. Removing these noise components helps to improve the denoising effect. In this way, the smoothed spectral matrix D3A is obtained i .

[0097] This spectral smoothing process can remove noise while retaining the key part of the spectral information, thereby enhancing the consistency of the denoised image in the spectral dimension. i The spectral characteristics of the image can be retained during the denoising process, making the denoised image more realistic in visual effect and maintaining the essential characteristics of the multispectral image.

[0098] 2.4 Tensor Reconstruction

[0099] After completing the spatial difference and spectral smoothing processing, the processed data is then reconstructed to obtain a denoised similarity tensor. Specifically, the spatial difference image D1B i and D2B iAnd the smoothed spectral matrix D3A i Combined together, a denoised non-local similarity tensor is generated. This reconstruction process is the inverse process of the above decomposition operation, integrating the differential spatial information and the processed spectral information, so that the advantages of each part are fully utilized to achieve the denoising effect.

[0100] Step 3: Exploring global low-rank properties of images after non-local self-similar processing

[0101] 3.1 Three-Direction Logarithmic Truncated Tensor Nuclear Norm (3TLogTNN) Calculation

[0102] In order to further improve the denoising effect of the image, a three-directional logarithmic truncated tensor nuclear norm (3TLogTNN) method is introduced in this embodiment to mine the global low-rank features of the image. In natural images, global low-rank features are usually manifested as redundancy and consistency under different modes, which are of great value in denoising. By introducing 3TLogTNN, the overall structural information of the image can be maintained while eliminating noise, thereby enhancing the denoising effect.

[0103] Specifically, we first perform TLogMNN calculation on the mode-k slice of the image tensor. The definition of TLogMNN is as follows:

[0104]

[0105] Among them, σ i represents the i-th singular value, ∈ is a small constant used to avoid taking the logarithm of zero value, usually a very small positive number is selected to ensure the stability of calculation, and m is the number of singular values. By performing logarithmic truncation operation on the singular values ​​of each modality one by one, TLogMNN can effectively retain the main singular values ​​of the image, thereby achieving noise suppression.

[0106] After performing TLogMNN calculations on the three modes of the image tensor, the resulting values ​​are added together and defined as 3TLogTNN. The process can be described as:

[0107] 3TLogTNN=TLogMNN(X (1) )+TLogMNN(X (2) )+TLogMNN(X (3) )

[0108] Among them, X (k)Represents the slice of the image tensor X in the kth mode (k = 1, 2, 3). This method can effectively retain the image details while suppressing noise through truncation and logarithmic operations. By selectively retaining the singular values ​​of each mode, 3TLogTNN ensures the global consistency of the image during the denoising process, so that the image maintains a high degree of clarity and detail integrity visually.

[0109] 3.2 Global Low-Rank Exploration

[0110] After completing the 3TLogTNN calculation, this step further explores the global low-rank characteristics of the image to preserve the overall structural information of the image. The global low-rank feature refers to the uniformity of the image under multimodality, which is common in natural images. By calculating 3TLogTNN, not only can the random noise of the image be removed, but also the low-rank structure of the image in a large range can be preserved, ensuring that the denoising effect covers the global information of the image.

[0111] The significance of global low-rank exploration is that the denoising process is extended to the entire image, rather than limited to local areas. This processing method can evenly remove noise in different areas, making the final denoised image more natural in both details and overall structure. In addition, the 3TLogTNN method avoids frequent tensor expansion operations, which makes the calculation of global low-rank exploration more efficient while reducing the loss of image details.

[0112] By combining the calculation of 3TLogTNN with the exploration of global low rank, it is possible to retain the global structural features of the image while removing noise, achieving a more efficient and accurate image denoising effect. This method is suitable for denoising complex data such as multispectral and hyperspectral images, and provides better image data support for high-precision detection in AOI systems.

[0113] Step 4: Build an image denoising model

[0114] In this step, the final image denoising model is constructed by combining the non-local two-factor regularizer and the three-directional logarithmic truncated tensor nuclear norm (3TLogTNN) regularizer. The optimization objective function of the model is designed as follows:

[0115]

[0116] The definitions of variables and parameters are as follows:

[0117] Y: collected original image data;

[0118] The denoised image represents the noise-free image finally restored by the model;

[0119] S: sparse noise term, used to compensate for the residual high-frequency noise components in the image, especially suitable for removing isolated outliers or noise spots;

[0120] G i =B i ×3A i : Reconstruction of the non-local similarity tensor, representing the representation of the image patch obtained by modulo three tensor decomposition, where B i is the spatial factor tensor, A i is the spectral matrix;

[0121] The reconstruction result G of all similar blocks i The complete denoised image is obtained by combining them.

[0122] The various terms of the optimization objective function are used to control the image restoration quality and noise suppression effect in the denoising process. Specifically:

[0123] Represents the reconstruction error term between the denoised image and the original image. The purpose of this term is to ensure that the denoised image The combination of and the sparse noise term S is as close to the original image Y as possible, thereby reducing the loss of image information.

[0124] The three-directional logarithmic truncated tensor nuclear norm (3TLogTNN) regularization term is used to maintain the global low-rank characteristics of the image during the denoising process, thereby enhancing the smoothness and consistency of the overall image structure.

[0125] Used to constrain the spectral matrix A i and its smoothed version The difference between them is calculated to ensure that the smoothness and continuity of the spectral information are preserved during denoising.

[0126] For the spectral matrix A i The smoothing constraint term after the difference is performed to reduce the fluctuation in the spectral dimension and suppress the interference of noise on the spectral information.

[0127] γ‖S‖1: Sparse regularization term, obtained by The norm constrains the sparse noise term S, making it tend to be sparsely distributed, which is particularly suitable for suppressing isolated high-frequency noise.

[0128] The parameters ρ, μ, and γ in the optimization objective function are the weight coefficients of the regularization term, which are used to achieve a balance between noise suppression and detail retention in the denoising process. By reasonably adjusting these parameters, the sensitivity of the denoising model to details and the degree of noise suppression can be controlled to achieve the best denoising effect. Through the above optimization process, image data with excellent denoising effect and complete detail retention is finally obtained, thereby significantly improving the image processing quality and detection accuracy of the AOI system.

[0129] The above description is only a preferred embodiment of the present invention, so all equivalent changes or modifications made according to the structure, characteristics and principles described in the scope of the patent application of the present invention are included in the scope of the patent application of the present invention.

Claims

1. An image denoising method for an automatic optical inspection system in the electronics industry, characterized in that: The following steps are involved: 1) Divide the collected images into blocks and group them according to similarity, including the following sub-steps: Dividing the acquired image into a number of sub-regions to form a plurality of three-dimensional image blocks; Calculate the similarity between each three-dimensional image block and other image blocks by using a similarity measurement method; According to the similarity, similar image blocks are grouped to form several similar groups containing multiple image blocks; Construct a three-dimensional tensor for each similar group of image patches; 2) Low-rank tensor modeling of similar blocks includes the following sub-steps: The similarity group tensor G i Decomposed into spatial factor tensor B i and the spectral matrix A i , where i is the similarity group index currently being processed, indicating the i-th similarity group; For the spatial factor tensor B i Processing to extract spatial features of the image; For the spectral matrix A i Processing is performed to smooth spectral features; Synthesize the processed spatial factor tensor B i and the spectral matrix A i , generate the denoised similar tensor G i ; 3) Perform global low-rank exploration on the denoised image, including the following sub-steps: Compute multimodal low-rank properties of images; Use global regularization methods to explore the global low-rank nature of images; 4) Construct an image denoising model to achieve noise suppression and detail retention by optimizing the objective function, where the optimization objective function is: in: Y: collected original image data; Denoised image; S: sparse noise term; G i : The i-th similarity group tensor obtained by tensor decomposition; B i : The spatial factor tensor of the i-th similarity group; A i : The spectral matrix of the i-th similarity group; Spectral matrix A i A smoothed version of D3A i : Spectral matrix A i The difference representation of ρ: regularization parameter; μ: spectral smoothing regularization parameter; γ: sparse regularization parameter; The square of the Frobenius norm; ||·|| 3DTLogTNN : three-directional logarithmic truncated tensor nuclear norm; ||·||1: l1 norm; n: Total number of similarity groups.

2. The image denoising method for use in an automatic optical inspection system in the electronics industry as claimed in claim 1, characterized in that: The non-local two-factor regularizer includes the spatial difference image D1B i and D2B i constraints.

3. The image denoising method for use in an automatic optical inspection system in the electronics industry as claimed in claim 1, characterized in that: The spectral smoothing includes the following steps: i Perform singular value decomposition.

4. The image denoising method for use in an automatic optical inspection system in the electronics industry as claimed in claim 1, characterized in that: The calculation of the three-way logarithmic truncated tensor nuclear norm (3TLogTNN) includes the TLogMNN calculation of the mode-k slice of the tensor, and its formula is as follows: in: X: input tensor slice; σ i : the i-th singular value of the input tensor slice X; ∈: a very small positive number; m: The number of singular values ​​retained.

5. The image denoising method for use in an automatic optical inspection system in the electronics industry as claimed in claim 1, characterized in that: The objective function of the image denoising model includes the following constraints: in: G i : represents the low-rank tensor representation of the i-th similar group obtained by tensor decomposition; B i : The spatial factor tensor of the ith similarity block; A i : The spectral matrix of the ith similarity block; ×3: represents tensor multiplication along the third mode (channel direction); Denoised image; n: The total number of similar chunks.

Images