A multi-source image-oriented automatic acquisition and quality enhancement method and system

Abstract

This invention provides an automated acquisition and quality enhancement method and system for multi-source images, belonging to the field of image enhancement technology. This invention provides a precise quantitative basis for global reliability and local sharpness of multi-source images through a dual-track quality evaluation system. First, the RANSAC algorithm is weighted using quality evaluation information, improving the robustness and alignment accuracy of feature registration in scenarios with uneven image quality. By combining a decision mechanism of global scoring and local variance, the clearest and most reliable detail source among all images is adaptively selected pixel-by-pixel, thereby maximizing the preservation of advantageous textures while improving the sharpness of individual images. When optimizing the low-frequency contours of each image, a common reference template is generated through nonlinear weighting based on global scoring and adaptively fused, improving the overall contrast and sense of depth of the image and suppressing noise. Finally, a fully automated closed-loop process for multi-source image acquisition and enhancement is achieved.

Description

Technical Field

[0001] This invention relates to the field of image enhancement technology, and in particular to an automated acquisition and quality enhancement method and system for multi-source images. Background Technology

[0002] In the automated processing of multi-source images, improving the overall quality of a collection of images of the same scene acquired under different sensor conditions presents a significant challenge. Existing methods typically denoise or enhance individual images independently, failing to fully utilize the complementary information between multiple source images. When attempting to use multiple images for collaborative enhancement, the primary difficulty lies in the reliability of image registration. Traditional feature-point matching methods, when image quality is uneven, blurry, or noisy, can lead to mismatches that severely affect registration accuracy, thereby undermining the foundation for subsequent collaborative processing. Secondly, there is a lack of effective decision-making mechanisms regarding how to utilize registered multiple images to optimize the quality of each individual image. Simple image averaging or selection can provide partial improvement, but it cannot intelligently identify and introduce clearer and more reliable information from other images to replace weaker parts in the target image at both the global contour and local detail levels. This results in limited enhancement effects, and the overall contrast, sharpness, and detail richness of the image cannot be systematically and adaptively improved comprehensively.

[0003] Therefore, it is necessary to provide an automated acquisition and quality enhancement method and system for multi-source images to solve the above-mentioned technical problems. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides an automated acquisition and quality enhancement method and system for multi-source images, achieving the beneficial effect of effectively automating the acquisition and quality enhancement of multi-source images.

[0005] This invention provides an automated acquisition and quality enhancement method for multi-source images, comprising:

[0006] S1: Simultaneously acquire initial images from multiple sources, extract local spatial features and global statistical features of each initial image, and calculate the local quality score map and global quality score of each initial image in combination with the acquisition parameters.

[0007] S2: Using the initial image with the highest global quality score as a benchmark, the RANSAC algorithm, which uses feature point matching and local quality score map weighting, is used to register each initial image to obtain a set of registered images.

[0008] S3: Perform multi-level discrete wavelet decomposition on each registered image to obtain the low-frequency sub-band coefficient matrix and high-frequency sub-band coefficient matrix of each registered image;

[0009] S4: Based on the global quality score and the local variance of the coefficient matrix of each high-frequency sub-band, generate the fusion index matrix of each high-frequency sub-band, generate the high-frequency coefficient selection mapping map of each high-frequency sub-band based on the fusion index matrix, and generate the enhanced high-frequency sub-band coefficient matrix of each high-frequency sub-band based on the high-frequency coefficient selection mapping map.

[0010] S5: Based on the global quality score, perform pixel-by-pixel weighted averaging on the low-frequency subband coefficient matrix to generate a low-frequency coefficient reference matrix. Then, perform weighted fusion on the low-frequency subband coefficient matrix and the low-frequency coefficient reference matrix to generate an enhanced low-frequency subband coefficient matrix.

[0011] S6: Perform inverse wavelet transform reconstruction on the enhanced high-frequency subband coefficient matrix and enhanced low-frequency subband coefficient matrix of each registered image to obtain a set of multi-source enhanced images.

[0012] Preferably, in step S1, the acquisition parameters include the analog gain value and exposure time at the time of each initial image acquisition.

[0013] Preferably, in step S1, in the calculation of the global quality score and the local quality score map, the simulated gain value is used to negatively correct the score, and the exposure time is used to correct the score through a nonlinear mapping relationship.

[0014] Preferably, in step S1, the local spatial features include the local grayscale contrast and local high-frequency detail intensity within a neighborhood window of a preset size centered on each pixel, and the global statistical features include the grayscale standard deviation and high-frequency detail energy value of each initial image.

[0015] Preferably, in step S2, the RANSAC algorithm with local quality score map weighting uses the maximum value of the weighted inliers as the criterion for selecting the optimal registration transformation matrix.

[0016] Preferably, in step S2, the weighting method in the local quality score map weighted RANSAC algorithm is as follows:

[0017] For each pair of corresponding feature points obtained through feature point matching, the weight is defined as the product of the values ​​of the two feature points at their corresponding positions in the local quality score map of their respective initial images.

[0018] Preferably, in step S4, the step of generating the fusion index matrix includes:

[0019] The local variance matrix of each high-frequency subband coefficient matrix is ​​multiplied pixel by pixel with the corresponding global quality score.

[0020] Preferably, in step S5, the low-frequency sub-band coefficient matrix is ​​weighted pixel-by-pixel based on the global quality score to generate a low-frequency coefficient reference matrix, including:

[0021] The weight of each pixel in the low-frequency subband coefficient matrix is ​​proportional to the square of the global quality score of the corresponding initial image.

[0022] Preferably, in step S5, the weighted fusion of the low-frequency subband coefficient matrix and the low-frequency coefficient reference matrix is ​​achieved by linearly combining the low-frequency subband coefficient matrix and the low-frequency coefficient reference matrix, wherein the fusion weight coefficient of the low-frequency coefficient reference matrix is ​​between 0.2 and 0.8.

[0023] This invention provides an automated acquisition and quality enhancement system for multi-source images, comprising:

[0024] The acquisition and quality assessment module is used to simultaneously acquire initial images from multiple sources, extract the local spatial features and global statistical features of each initial image, and calculate the local quality score map and global quality score of each initial image in combination with the acquisition parameters.

[0025] The quality-weighted registration module is used to register each initial image with the highest global quality score as a reference, and to obtain a set of registered images by using the RANSAC algorithm that combines feature point matching and local quality score map weighting.

[0026] The multi-scale decomposition module is used to perform multi-level discrete wavelet decomposition on each registered image to obtain the low-frequency subband coefficient matrix and the high-frequency subband coefficient matrix of each registered image.

[0027] The high-frequency decision fusion module is used to generate a fusion index matrix for each high-frequency sub-band based on the global quality score and the local variance of the coefficient matrix of each high-frequency sub-band, generate a high-frequency coefficient selection mapping map for each high-frequency sub-band based on the fusion index matrix, and generate an enhanced high-frequency sub-band coefficient matrix for each high-frequency sub-band based on the high-frequency coefficient selection mapping map.

[0028] The low-frequency optimization fusion module is used to perform pixel-by-pixel weighted averaging of the low-frequency subband coefficient matrix based on the global quality score to generate a low-frequency coefficient reference matrix, and to perform weighted fusion of the low-frequency subband coefficient matrix and the low-frequency coefficient reference matrix to generate an enhanced low-frequency subband coefficient matrix.

[0029] The reconstruction output module is used to perform inverse wavelet transform reconstruction on the enhanced high-frequency subband coefficient matrix and enhanced low-frequency subband coefficient matrix of each registered image to obtain a set of multi-source enhanced images.

[0030] Compared with related technologies, the automated acquisition and quality enhancement method and system for multi-source images provided by this invention has the following beneficial effects:

[0031] This invention provides a precise quantitative foundation for global reliability and local sharpness of multi-source images through synchronous acquisition and dual-track quality evaluation. First, the quality evaluation information is used to weight the RANSAC algorithm, significantly improving the robustness and geometric alignment accuracy of feature registration in scenarios with uneven image quality, ensuring the reliability of multi-source information comparison. When reconstructing high-frequency details of each image, a decision mechanism combining global scoring and local variance adaptively selects the clearest and most reliable detail source from all images pixel by pixel, thereby maximizing the preservation of superior textures while improving the sharpness of individual images. When optimizing low-frequency contours of each image, a common reference template is generated through nonlinear weighting based on global scoring and adaptively fused, effectively improving the overall contrast and layering of the image while suppressing noise. Finally, a fully automated closed-loop process for multi-source image acquisition and enhancement, from acquisition, evaluation, precise registration to collaborative enhancement, is achieved. Attached Figure Description

[0032] Figure 1 This is a flowchart of an automated acquisition and quality enhancement method for multi-source images according to the present invention;

[0033] Figure 2 This is a block diagram of an automated acquisition and quality enhancement system for multi-source images according to the present invention. Detailed Implementation

[0034] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the drawings, not all structures. Moreover, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0035] It should also be noted that, for ease of description, the accompanying drawings show only the parts relevant to the invention and not all of them. Before discussing exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the operations (or steps) as sequential processes, many of the operations can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations can be rearranged. The process can be terminated when its operation is completed, but it may also have additional steps not included in the drawings. The process may correspond to a method, function, procedure, subroutine, subroutine, etc.

[0036] Example 1

[0037] An automated acquisition and quality enhancement method for multi-source images, in the specific implementation process, such as... Figure 1 As shown, a flowchart of an automated acquisition and quality enhancement method for multi-source images according to the present invention is illustrated, including:

[0038] Step S1: Simultaneously acquire multiple initial images from different sources, extract the local spatial features and global statistical features of each initial image, and calculate the local quality score map and global quality score of each initial image in combination with the acquisition parameters.

[0039] Specifically, in step S1, the acquisition parameters include the analog gain value and exposure time at the time of each initial image acquisition.

[0040] Specifically, in step S1, in the calculation of the global quality score and the local quality score map, the simulated gain value is used to negatively correct the score, and the exposure time is used to correct the score through a nonlinear mapping relationship.

[0041] Specifically, in step S1, the local spatial features include the local grayscale contrast and local high-frequency detail intensity within a neighborhood window of a preset size centered on each pixel, and the global statistical features include the grayscale standard deviation and high-frequency detail energy value of each initial image.

[0042] In the specific implementation process, firstly, at least two image sensors are triggered by a hardware synchronization signal to expose the same scene at the same physical moment, acquiring multiple initial images from multiple sources. The analog gain value and exposure time of each initial image are recorded. Next, for each initial image, the standard deviation of the grayscale values ​​of all pixels in the entire image is calculated as the grayscale standard deviation in the global statistical features. Simultaneously, each initial image undergoes high-pass filtering, and the sum of the squares of all pixel values ​​in the high-frequency components of the filtered output image is used as the high-frequency detail energy value in the global statistical features. Then, a local quality score map is generated. For example, for each initial image, a local quality score map is generated centered on each pixel. A square neighborhood window with a side length of 15 pixels is selected. The standard deviation of the gray values ​​of all pixels within this square neighborhood window is calculated. This standard deviation is used as the local gray-level contrast of the center pixel. Simultaneously, from the high-frequency component image of the whole image obtained by the high-pass filtering, the same neighborhood window containing the same center pixel is extracted. The sum of squares of the high-frequency component values ​​within this neighborhood window is calculated. This sum of squares is used as the local high-frequency detail intensity of the center pixel. This process is repeated for all pixels in the image to complete the calculation of all local spatial features. Then, the score is calculated. The global quality score is calculated, for example, by first normalizing the standard deviation of the gray levels of the entire image and the high-frequency detail energy values ​​to 0. The system first calculates the numerical range up to 1, then adds the two normalized values ​​with a preset fixed weight ratio to obtain the initial score. Next, it performs corrections based on the collected parameters. The first step is gain correction. For example, a baseline gain of 1 and a corresponding deduction coefficient of 0.05 are set. The ratio of the actual gain value to the baseline gain is calculated as the gain multiplier. This multiplier is multiplied by the deduction coefficient to obtain the negative reduction, which is subtracted from the initial score. For example, if the actual gain is 4, the calculated gain multiplier is 4, resulting in a negative reduction of 0.2. Finally, 0.2 is subtracted from the initial score to obtain the gain-corrected score. The second step is exposure correction. For example, the ideal exposure time for the current scene is preset to 10 milliseconds, and a maximum exposure time is set... The allowable deviation ratio is 1. The absolute value of the difference between the actual exposure time and the ideal value is calculated and then divided by the ideal value to obtain the relative deviation. If the relative deviation exceeds 1, it is taken as 1. Then, the nonlinear attenuation factor is calculated, which is the sum of the value 1 divided by 1 and the relative deviation. Finally, the score after gain correction is multiplied by the attenuation factor to obtain the final global quality score. The calculation of the local quality score map uses the same indicators, weights and correction rules as the calculation of the global quality score. The local grayscale contrast and local high-frequency detail intensity are calculated and corrected pixel by pixel, generating a local quality score value for each pixel. The score values ​​of all pixels are arranged in their original positions to form the local quality score map of the image.

[0043] Step S2: Using the initial image with the highest global quality score as a benchmark, register each initial image using the RANSAC algorithm that combines feature point matching and local quality score map weighting to obtain a set of registered images.

[0044] Specifically, in step S2, the local quality score map weighted RANSAC algorithm uses the maximum value of the weighted inliers as the criterion for selecting the optimal registration transformation matrix.

[0045] Specifically, in step S2, the weighting method in the local quality score map weighted RANSAC algorithm is as follows:

[0046] For each pair of corresponding feature points obtained through feature point matching, the weight is defined as the product of the values ​​of the two feature points at their corresponding positions in the local quality score map of their respective initial images.

[0047] In the specific implementation process, the image with the highest global quality score is selected as the reference image for geometric registration. If the global quality scores are the same, one of them is arbitrarily designated as the reference image. Next, on the reference image and each initial image to be registered, the scale-invariant feature transform algorithm is used to extract key points and calculate their feature descriptors. Then, the feature descriptors between the two initial images are matched using nearest neighbor search, establishing a pair of corresponding feature points represented by coordinate values ​​for each pair of images. Afterwards, a local quality score map-weighted random sampling consensus algorithm is executed. The specific implementation process of this algorithm is as follows: for each pair of corresponding feature points between the reference image and any image to be registered, the weight is uniquely defined as the product of the score values ​​recorded at the corresponding pixel coordinate positions in the local quality score map of the two feature points in that pair, and then the calculation begins. The iterative process of the method involves randomly selecting four pairs of corresponding feature points in each iteration to calculate a candidate perspective transformation matrix model. This candidate perspective transformation matrix model is then used to calculate the projection error of all corresponding feature point pairs. Feature point pairs with projection errors less than a preset threshold are identified as interior points of the current perspective transformation matrix model. The sum of all weight values ​​corresponding to these interior points is calculated to obtain the weighted interior point count for this iteration. After completing a preset number of iterations, the perspective transformation matrix model with the largest weighted interior point count is selected from all candidate perspective transformation matrix models, and its corresponding perspective transformation matrix is ​​ultimately determined as the optimal registration transformation matrix from the image to be registered to the reference image. Finally, this optimal transformation matrix is ​​used to perform a resampling geometric transformation on the image to be registered, aligning its pixel coordinate system with the reference image. This process is repeated for all images to be registered, and a set of registered images in which all initial images are spatially aligned is output.

[0048] Step S3: Perform multi-level discrete wavelet decomposition on each registered image to obtain the low-frequency subband coefficient matrix and high-frequency subband coefficient matrix of each registered image.

[0049] In a specific implementation, for example, a three-level discrete wavelet decomposition is first performed on each registered image in the registered image set, using the Daubechies 4 wavelet as the wavelet basis function. The decomposition process is as follows: the input two-dimensional registered image is first subjected to a one-dimensional discrete wavelet transform in the row direction, and then a one-dimensional discrete wavelet transform in the column direction. After each two-dimensional transform, the image is decomposed into four sub-band coefficient matrices with halved size, including a low-frequency approximate sub-band coefficient matrix containing the main contour information, a high-frequency horizontal detail sub-band coefficient matrix containing horizontal details, a high-frequency vertical detail sub-band coefficient matrix containing vertical details, and a high-frequency diagonal detail sub-band coefficient matrix containing diagonal details. After completing the first level of decomposition, the obtained low-frequency approximate sub-band coefficient matrix is ​​used as input, and the above two-dimensional discrete wavelet decomposition is repeated. In the wavelet transform process, a second-level decomposition is performed, resulting in four sub-band coefficient matrices of the second level with their size halved again. Next, the low-frequency approximate sub-band coefficient matrices obtained from the second-level decomposition are used as input for a third-level decomposition, resulting in four sub-band coefficient matrices of the third level with the smallest size. After three levels of decomposition, for each registered image, the low-frequency sub-band coefficient matrix of the deepest decomposition layer and multiple high-frequency sub-band coefficient matrices are finally obtained. Specifically, this includes one low-frequency approximate sub-band coefficient matrix of the third level, and three high-frequency detail sub-band coefficient matrices contained in the first, second, and third levels, for a total of ten coefficient matrices, which serve as inputs for subsequent steps.

[0050] Step S4: Based on the global quality score and the local variance of the coefficient matrix of each high-frequency sub-band, generate the fusion index matrix of each high-frequency sub-band; based on the fusion index matrix, generate the high-frequency coefficient selection mapping map of each high-frequency sub-band; based on the high-frequency coefficient selection mapping map, generate the enhanced high-frequency sub-band coefficient matrix of each high-frequency sub-band.

[0051] Specifically, in step S4, the steps for generating the fusion index matrix include:

[0052] The local variance matrix of each high-frequency subband coefficient matrix is ​​multiplied pixel by pixel with the corresponding global quality score.

[0053] In the specific implementation process, for each image in the registered image set, after undergoing multi-layer discrete wavelet decomposition, a high-frequency sub-band of a specific decomposition layer and a specific direction is obtained. Then, a high-frequency sub-band coefficient matrix of the same size is acquired for all registered images in that specific direction. Next, the local variance matrix corresponding to the high-frequency coefficient matrix of each registered image is calculated. For example, the variance of all high-frequency coefficient values ​​within a 5-pixel × 5-pixel square neighborhood window surrounding each pixel in the high-frequency sub-band coefficient matrix is ​​calculated, and this variance value is assigned to the center pixel. After traversing all pixels, a local variance matrix with the same size as the high-frequency sub-band coefficient matrix is ​​obtained. The difference matrix is ​​then processed. Next, the core operation of generating the fusion index matrix is ​​performed: multiplying each pixel value in the local variance matrix of each registered image with the global quality score of the corresponding initial image. This generates a fusion index matrix for each registered image with the same size as the high-frequency subband coefficient matrix of that decomposition layer and a specific direction. Subsequently, a high-frequency coefficient selection map is generated based on the fusion index matrices of all high-frequency subband coefficient matrices. Specifically, the generation rule is to compare the fusion index values ​​of each fusion index matrix at the same pixel coordinate position of these fusion index matrices with the same decomposition layer, direction, and size. The image number corresponding to the largest fusion index value is selected, and this image number is recorded as the value of the high-frequency coefficient selection map at that pixel position. After traversing all pixel positions, a high-frequency coefficient selection map with the same size as the high-frequency subband coefficient matrix is ​​finally obtained, but each pixel position stores a registration image number. Finally, based on this high-frequency coefficient selection map, an enhanced high-frequency subband coefficient matrix is ​​generated for each registration image. The specific generation rule is that for a registration image to which the enhanced high-frequency subband coefficient matrix is ​​to be generated, at each pixel position of the map, the recorded registration image number is read; if the number points to the registration image... For a given image, the coefficient value at the corresponding position in its own high-frequency subband coefficient matrix is ​​used as the value of the enhanced high-frequency subband coefficient matrix at that position. If the number points to another registered image, the coefficient value at the corresponding position in the high-frequency subband coefficient matrix of that registered image is taken out and used as the value of the enhanced high-frequency subband coefficient matrix at that position. By traversing all pixel positions, a brand-new enhanced high-frequency subband coefficient matrix that integrates the best high-frequency information from multiple sources is generated for the registered image. The above process is repeated for each decomposition layer and each direction of the high-frequency subband, thereby generating a complete set of enhanced high-frequency subband coefficient matrices for each registered image.

[0054] Step S5: Perform pixel-by-pixel weighted averaging on the low-frequency subband coefficient matrix based on the global quality score to generate a low-frequency coefficient reference matrix. Then, perform weighted fusion of the low-frequency subband coefficient matrix and the low-frequency coefficient reference matrix to generate an enhanced low-frequency subband coefficient matrix.

[0055] Specifically, in step S5, the low-frequency sub-band coefficient matrix is ​​weighted pixel-by-pixel based on the global quality score to generate a low-frequency coefficient reference matrix, including:

[0056] The weight of each pixel in the low-frequency subband coefficient matrix is ​​proportional to the square of the global quality score of the corresponding initial image.

[0057] Specifically, in step S5, the weighted fusion of the low-frequency sub-band coefficient matrix and the low-frequency coefficient reference matrix is ​​achieved by linearly combining the low-frequency sub-band coefficient matrix and the low-frequency coefficient reference matrix, wherein the fusion weight coefficient of the low-frequency coefficient reference matrix is ​​between 0.2 and 0.8.

[0058] In the specific implementation process, firstly, the low-frequency subband coefficient matrix generated at the deepest layer after multi-level discrete wavelet decomposition of all registered images is obtained, and the global quality score of the initial image corresponding to each registered image is read. Next, the square of the global quality score of the initial image corresponding to each registered image is calculated, and then the squares of all global quality scores are summed. The square of the global quality score of the corresponding initial image is divided by this sum to calculate a normalized weight for each registered image. This normalized weight is proportional to the square of the global quality score of the initial image corresponding to that registered image, and the sum of the normalized weights of all registered images is one. Subsequently, to output each pixel position of the low-frequency coefficient reference matrix, the coefficient value at the same coordinate position is extracted from the low-frequency subband coefficient matrix of all registered images. Each coefficient value is multiplied by the normalized weight of its corresponding registered image, and then all weighted coefficient values ​​are summed. The resulting sum is the value of the low-frequency coefficient reference matrix at that pixel position. After traversing all pixel positions, a complete low-frequency coefficient reference matrix is ​​generated. Then, a specific fusion weight coefficient value is set, which is in the range of 0.2 to 0.8. For each registered image to be enhanced, at each pixel position of its low-frequency subband coefficient matrix, its original coefficient value and the value of the low-frequency coefficient reference matrix at the same position are read and linearly combined. For example, the coefficient value of each pixel position of the low-frequency subband coefficient matrix is ​​multiplied by the difference obtained by subtracting the fusion weight coefficient to obtain the first product. Then, the coefficient value of the corresponding pixel position of the low-frequency coefficient reference matrix is ​​multiplied by the fusion weight coefficient to obtain the second product. Finally, the first product and the second product are added together. The resulting value is the value of the low-frequency subband coefficient matrix of the registered image after enhancement at this pixel position. Traversing all pixel positions of the low-frequency subband coefficient matrix completes the generation of its enhanced low-frequency subband coefficient matrix, thereby generating its independent enhanced low-frequency subband coefficient matrix for each registered image.

[0059] Step S6: Perform inverse wavelet transform reconstruction on the enhanced high-frequency subband coefficient matrix and enhanced low-frequency subband coefficient matrix of each registered image to obtain a set of multi-source enhanced images.

[0060] In the specific implementation process, firstly, the enhanced low-frequency subband coefficient matrix corresponding to the registered image and the three enhanced high-frequency subband coefficient matrices belonging to the deepest decomposition are used as input. Using the same wavelet basis function Daubechies 4 wavelet as during decomposition, a single-level inverse two-dimensional discrete wavelet transform is performed on the deepest enhanced low-frequency subband coefficient matrix and the three enhanced high-frequency subband coefficient matrices. After the transform, a second-level low-frequency coefficient recovery matrix with its size doubled is output. Then, the second-level low-frequency coefficient recovery matrix and the three enhanced high-frequency subband coefficient matrices corresponding to the second-level decomposition are used as input, and the same single-level inverse two-dimensional discrete wavelet transform is performed again to obtain a first-level low-frequency coefficient recovery matrix with its size doubled again. Next, this first-level low-frequency coefficient recovery matrix and the three enhanced high-frequency subband coefficient matrices corresponding to the first-level decomposition are used as input for a final single-level inverse two-dimensional discrete wavelet transform. Finally, an enhanced image reconstructed to the original spatial domain size is obtained, thereby generating a corresponding final quality enhanced image for each initial image. The set of all final quality enhanced images constitutes a multi-source enhanced image set.

[0061] The working principle of the automated acquisition and quality enhancement method for multi-source images provided by this invention is as follows:

[0062] First, through synchronous acquisition and dual-track quality evaluation, a global quality score reflecting overall reliability and a local quality score map revealing the spatial distribution of local sharpness are simultaneously generated for each image, thereby quantifying its information reliability. Then, in the image registration stage, the local quality score map is used to assign differentiated weights to feature point matching, guiding the RANSAC algorithm to solve the transformation model by maximizing the number of weighted inliers, thus achieving high-precision geometric alignment results even when the source images have uneven quality. Based on this, multi-scale wavelet decomposition is performed on the registered images, and during high-frequency detail reconstruction, an augmentation map is independently generated for each image. The principle of the enhancement coefficients is to combine the local activity of the high-frequency coefficients of each source image with its global quality score, make pixel-by-pixel decisions, and select the clearest and most reliable source of detail among all images. For low-frequency contours, the principle is to first synthesize a high-quality common reference background based on the nonlinear weight of the global quality score of each image when generating enhancement coefficients for each image, and then adaptively linearly combine the low-frequency information of each image with the reference background to optimize the overall contrast and suppress noise. Finally, the inverse wavelet transform is used to independently reconstruct an enhanced version of the image for each input image, which incorporates the complementary advantages of multiple sources and has significantly improved quality.

[0063] Example 2

[0064] An automated acquisition and quality enhancement system for multi-source images, in its specific implementation, such as... Figure 2As shown, it illustrates a modular structure diagram of an automated acquisition and quality enhancement system for multi-source images according to the present invention, including:

[0065] The acquisition and quality evaluation module 100 is used to simultaneously acquire multiple initial images from multiple sources, extract the local spatial features and global statistical features of each initial image, and calculate the local quality score map and global quality score of each initial image in combination with the acquisition parameters.

[0066] The quality-weighted registration module 200 is used to register each initial image with the highest global quality score as a reference, and obtain a set of registered images by using the RANSAC algorithm that combines feature point matching and local quality score map weighting.

[0067] The multi-scale decomposition module 300 is used to perform multi-level discrete wavelet decomposition on each registered image to obtain the low-frequency subband coefficient matrix and the high-frequency subband coefficient matrix of each registered image.

[0068] The high-frequency decision fusion module 400 is used to generate a fusion index matrix for each high-frequency sub-band based on the global quality score and the local variance of the coefficient matrix of each high-frequency sub-band, generate a high-frequency coefficient selection mapping map for each high-frequency sub-band based on the fusion index matrix, and generate an enhanced high-frequency sub-band coefficient matrix for each high-frequency sub-band based on the high-frequency coefficient selection mapping map.

[0069] The low-frequency optimization fusion module 500 is used to perform pixel-by-pixel weighted averaging of the low-frequency subband coefficient matrix based on the global quality score to generate a low-frequency coefficient reference matrix, and to perform weighted fusion of the low-frequency subband coefficient matrix and the low-frequency reference matrix to generate an enhanced low-frequency subband coefficient matrix.

[0070] The reconstruction output module 600 is used to perform inverse wavelet transform reconstruction on the enhanced high-frequency subband coefficient matrix and enhanced low-frequency subband coefficient matrix of each registered image to obtain a set of multi-source enhanced images.

[0071] The working principle of the automated acquisition and quality enhancement system for multi-source images provided by this invention is as follows:

[0072] This invention employs an acquisition and quality evaluation module 100 to synchronously acquire and calculate the global and local quality scores for each image, providing a quantitative quality basis for the system. Based on this, a quality-weighted registration module 200 uses the local quality score to weight the RANSAC algorithm, achieving high-precision image registration. A multi-scale decomposition module 300 then performs wavelet transform on the registered images, separating the low-frequency and high-frequency coefficient matrices. A high-frequency decision fusion module 400 compares the local variance of the high-frequency coefficients of each image and combines it with their global quality scores to independently decide for each image and select the optimal source of high-frequency details, generating its enhanced high-frequency coefficients. A low-frequency optimization fusion module 500 generates a high-quality common low-frequency reference template based on the global quality score and guides the low-frequency coefficients of each image to optimize and fuse with it, improving the background and suppressing noise. Finally, a reconstruction output module 600 performs inverse wavelet transform on the independent enhanced high- and low-frequency coefficients of each image, reconstructing each significantly improved enhanced image, thus completing the entire automated enhancement process.

[0073] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0074] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-Erasable Programmable Read-Only Memory (EEPROM), compactdisc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium capable of carrying or storing data.

[0075] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

Claims

1. An automated acquisition and quality enhancement method for multi-source images, characterized in that, The enhancement method includes: S1: Simultaneously acquire initial images from multiple sources, extract local spatial features and global statistical features of each initial image, and calculate the local quality score map and global quality score of each initial image in combination with the acquisition parameters. S2: Using the initial image with the highest global quality score as a benchmark, the RANSAC algorithm, which uses feature point matching and local quality score map weighting, is used to register each initial image to obtain a set of registered images. S3: Perform multi-level discrete wavelet decomposition on each registered image to obtain the low-frequency sub-band coefficient matrix and high-frequency sub-band coefficient matrix of each registered image; S4: Based on the global quality score and the local variance of the coefficient matrix of each high-frequency sub-band, generate the fusion index matrix of each high-frequency sub-band, generate the high-frequency coefficient selection mapping map of each high-frequency sub-band based on the fusion index matrix, and generate the enhanced high-frequency sub-band coefficient matrix of each high-frequency sub-band based on the high-frequency coefficient selection mapping map. S5: Based on the global quality score, perform pixel-by-pixel weighted averaging on the low-frequency subband coefficient matrix to generate a low-frequency coefficient reference matrix. Then, perform weighted fusion on the low-frequency subband coefficient matrix and the low-frequency coefficient reference matrix to generate an enhanced low-frequency subband coefficient matrix. S6: Perform inverse wavelet transform reconstruction on the enhanced high-frequency subband coefficient matrix and enhanced low-frequency subband coefficient matrix of each registered image to obtain a set of multi-source enhanced images.

2. The automated acquisition and quality enhancement method for multi-source images according to claim 1, characterized in that, In step S1, the acquisition parameters include the analog gain value and exposure time at the time of each initial image acquisition.

3. The automated acquisition and quality enhancement method for multi-source images according to claim 2, characterized in that, In step S1, in the calculation of the global quality score and the local quality score map, the simulated gain value is used to negatively correct the score, and the exposure time is used to correct the score through a nonlinear mapping relationship.

4. The automated acquisition and quality enhancement method for multi-source images according to claim 3, characterized in that, In step S1, the local spatial features include the local grayscale contrast and local high-frequency detail intensity within a neighborhood window of a preset size centered on each pixel, and the global statistical features include the grayscale standard deviation and high-frequency detail energy value of each initial image.

5. The automated acquisition and quality enhancement method for multi-source images according to claim 4, characterized in that, In step S2, the local quality score map weighted RANSAC algorithm uses the maximum number of weighted inliers as the criterion for selecting the optimal registration transformation matrix.

6. The automated acquisition and quality enhancement method for multi-source images according to claim 5, characterized in that, In step S2, the weighting method in the local quality score map weighted RANSAC algorithm is as follows: For each pair of corresponding feature points obtained through feature point matching, the weight is defined as the product of the values ​​of the two feature points at their corresponding positions in the local quality score map of their respective initial images.

7. The automated acquisition and quality enhancement method for multi-source images according to claim 6, characterized in that, In step S4, the steps for generating the fusion index matrix include: The local variance matrix of each high-frequency subband coefficient matrix is ​​multiplied pixel by pixel with the corresponding global quality score.

8. The automated acquisition and quality enhancement method for multi-source images according to claim 7, characterized in that, In step S5, a pixel-by-pixel weighted average is performed on the low-frequency subband coefficient matrix based on the global quality score to generate a low-frequency coefficient reference matrix, including: The weight of each pixel in the low-frequency subband coefficient matrix is ​​proportional to the square of the global quality score of the corresponding initial image.

9. The automated acquisition and quality enhancement method for multi-source images according to claim 8, characterized in that, In step S5, the low-frequency sub-band coefficient matrix and the low-frequency coefficient reference matrix are weighted and fused by linearly combining the low-frequency sub-band coefficient matrix and the low-frequency coefficient reference matrix. The fusion weight coefficient of the low-frequency coefficient reference matrix is ​​between 0.2 and 0.

8.

10. An automated acquisition and quality enhancement system for multi-source images, characterized in that, An automated acquisition and quality enhancement method for multi-source images, as described in any one of claims 1 to 9, wherein the enhancement system comprises: The acquisition and quality assessment module is used to simultaneously acquire initial images from multiple sources, extract the local spatial features and global statistical features of each initial image, and calculate the local quality score map and global quality score of each initial image in combination with the acquisition parameters. The quality-weighted registration module is used to register each initial image with the highest global quality score as a reference, and to obtain a set of registered images by using the RANSAC algorithm that combines feature point matching and local quality score map weighting. The multi-scale decomposition module is used to perform multi-level discrete wavelet decomposition on each registered image to obtain the low-frequency subband coefficient matrix and the high-frequency subband coefficient matrix of each registered image. The high-frequency decision fusion module is used to generate a fusion index matrix for each high-frequency sub-band based on the global quality score and the local variance of the coefficient matrix of each high-frequency sub-band, generate a high-frequency coefficient selection mapping map for each high-frequency sub-band based on the fusion index matrix, and generate an enhanced high-frequency sub-band coefficient matrix for each high-frequency sub-band based on the high-frequency coefficient selection mapping map. The low-frequency optimization fusion module is used to perform pixel-by-pixel weighted averaging of the low-frequency subband coefficient matrix based on the global quality score to generate a low-frequency coefficient reference matrix, and to perform weighted fusion of the low-frequency subband coefficient matrix and the low-frequency coefficient reference matrix to generate an enhanced low-frequency subband coefficient matrix. The reconstruction output module is used to perform inverse wavelet transform reconstruction on the enhanced high-frequency subband coefficient matrix and enhanced low-frequency subband coefficient matrix of each registered image to obtain a set of multi-source enhanced images.

Images