Machine vision image processing smoothing method and system

Abstract

The invention provides a machine vision image processing smoothing method and system. The method comprises the following steps: acquiring an original machine vision image, and performing preliminary noise reduction processing on the original machine vision image through a fast non-local mean algorithm and a box filter to acquire an image after preliminary noise reduction; performing end-to-end low-resolution image super-resolution reconstruction on the image subjected to preliminary noise reduction through a deep residual convolutional network to obtain a super-resolution reconstructed image; carrying out image edge analysis screening on the super-resolution reconstructed image to screen out an edge fuzzy gray level out-of-tolerance image; and carrying out edge blur elimination and gray level recovery on the edge blur gray level out-of-tolerance image by selecting an edge smoothing preserving model.

Description

Technical Field

[0001] The present invention relates to the technical field of high-precision remote sensing measurement image processing, and more specifically, to a machine vision image processing smoothing method and system. Background Art

[0002] At present, deep learning has become a development trend and research hotspot in the field of machine learning. The application of deep convolutional neural networks in image segmentation, image recognition, and autonomous driving has introduced new ideas for image denoising. Remote sensing images, medical images, astronomical images, synthetic aperture radar images, night vision images, and other images in different fields are interfered by the equipment itself and external signals during the imaging and transmission process, resulting in a large amount of noise in the image, and the image details are submerged in the noise; or the image resolution is reduced, which seriously affects the image quality, resulting in subsequent image segmentation, image recognition, etc. cannot be carried out smoothly; image denoising and image super-resolution reconstruction key image processing and other problems remain to be solved; therefore, it is necessary to propose an image processing smoothing method and system for machine vision to at least partially solve the problems existing in the prior art. Summary of the invention

[0003] A series of simplified concepts are introduced in the content of the invention, which will be further described in detail in the specific implementation method section; the content of the invention of the present invention does not mean to attempt to limit the key features and essential technical features of the technical solution claimed for protection, nor does it mean to attempt to determine the scope of protection of the technical solution claimed for protection.

[0004] To at least partially solve the above problems, the present invention provides a machine vision image processing smoothing method, comprising:

[0005] S100, obtaining a machine vision original image, performing preliminary noise reduction processing on the machine vision original image by using a fast non-local mean algorithm and a box filter, and obtaining a preliminary noise-reduced image;

[0006] S200, the image after preliminary denoising is subjected to end-to-end low-resolution image super-resolution reconstruction through a deep residual convolutional network to obtain a super-resolution reconstructed image;

[0007] S300, performing image edge analysis and screening on the super-resolution reconstructed image to screen out edge fuzzy grayscale super-poor images;

[0008] S400, for the edge blurred grayscale out-of-parity image, the edge-preserving smoothing model is selected to remove the edge blur and restore the grayscale.

[0009] Preferably, S100 includes:

[0010] S101, obtaining an original machine vision image, and performing a first denoising process on the original machine vision image using a fast non-local means algorithm to obtain a first denoising processed image;

[0011] S102: Perform preliminary denoising on the first denoised image by using a box filter to obtain a preliminary denoised image.

[0012] Preferably, S200 includes:

[0013] S201, construct a deep residual convolutional network through convolution propagation stacking;

[0014] S202, inputting the image after preliminary denoising into a deep residual convolutional network;

[0015] S203, performing end-to-end low-resolution image super-resolution reconstruction through a deep residual convolutional network to obtain a super-resolution reconstructed image.

[0016] Preferably, S300 includes:

[0017] S301, setting an image edge analysis screening model by identifying the image edge region;

[0018] S302, performing image edge analysis and screening on the image edge area of ​​the super-resolution reconstructed image through an image edge analysis and screening model to screen out edge fuzzy grayscale super-difference images.

[0019] Preferably, S400 includes:

[0020] S401, creating a selected edge-preserving smoothing model according to an edge-preserving uniform smoothing processing framework; inputting an edge-blurred grayscale super-difference image into the selected edge-preserving smoothing model;

[0021] S402, removing edge blur and restoring grayscale in the edge blur grayscale excess area by selecting an edge preserving smoothing model.

[0022] The present invention provides a machine vision image processing smoothing system, comprising:

[0023] The local mean box filter subsystem obtains the original machine vision image, performs preliminary noise reduction processing on the original machine vision image through the fast non-local mean algorithm and the box filter, and obtains the image after preliminary noise reduction;

[0024] Deep convolution image reconstruction subsystem: After preliminary denoising, the image is subjected to end-to-end low-resolution image super-resolution reconstruction through a deep residual convolutional network to obtain a super-resolution reconstructed image;

[0025] The edge analysis and screening subsystem performs image edge analysis and screening on the super-resolution reconstructed image to screen out images with blurred edges and grayscale out-of-parity;

[0026] The edge-preserving smoothing restoration subsystem selects an edge-preserving smoothing model to remove edge blur and restore grayscale for images with edge blur and grayscale out-of-parity.

[0027] Preferably, the local mean box filter subsystem includes:

[0028] A non-local mean fast calculation module is used to obtain a machine vision original image, and to perform a first denoising process on the machine vision original image by using a fast non-local mean algorithm to obtain a first denoising processed image;

[0029] The box filter denoising module performs preliminary denoising on the first denoising image through a box filter according to the first denoising image, so as to obtain a preliminary denoising image.

[0030] Preferably, the deep convolution image reconstruction subsystem includes:

[0031] The propagation stacking deep convolution module constructs a deep residual convolution network through convolution propagation stacking;

[0032] The residual convolutional network input module inputs the image after preliminary denoising into the deep residual convolutional network;

[0033] The end-to-end super-resolution reconstruction module performs end-to-end super-resolution reconstruction of low-resolution images through a deep residual convolutional network to obtain a super-resolution reconstructed image.

[0034] Preferably, the edge analysis screening subsystem comprises:

[0035] The edge region recognition module sets the image edge analysis and screening model by identifying the image edge region;

[0036] The out-of-tolerance analysis area screening module uses the image edge analysis screening model to perform image edge analysis screening on the image edge area of ​​the super-resolution reconstructed image to screen out edge blurred grayscale out-of-tolerance images.

[0037] Preferably, the edge-preserving smooth restoration subsystem includes:

[0038] The edge uniform smoothing model module creates a selected edge preserving smoothing model according to the edge uniform smoothing processing framework; the edge fuzzy grayscale super-difference image is input into the selected edge preserving smoothing model;

[0039] The edge blur out-of-tolerance restoration module selects an edge-preserving smoothing model to remove edge blur and restore the grayscale of the edge blur out-of-tolerance grayscale area.

[0040] Compared with the prior art, the present invention has at least the following beneficial effects:

[0041] The present invention discloses a machine vision image processing smoothing method and system, which obtains a machine vision original image, performs preliminary noise reduction processing on the machine vision original image through a fast non-local mean algorithm and a box filter, and obtains a preliminary noise-reduced image; performs end-to-end low-resolution image super-resolution reconstruction on the preliminary noise-reduced image through a deep residual convolutional network, and obtains a super-resolution reconstructed image; performs image edge analysis and screening on the super-resolution reconstructed image, and screens out edge blurred grayscale super-error images; for edge blurred grayscale super-error images, performs edge blur removal and grayscale restoration by selecting an edge-preserving smoothing model; makes full use of a large amount of redundant information existing in natural images, can better maintain the details and texture of the image, and can effectively prevent the occurrence of image malfunctions. The large receptive field super-resolution reconstruction network convolutional neural network model makes full use of the contextual information of a large number of samples, learns the implicit details and texture information in the samples, and combines the advantages of both aspects; the denoising method of the present invention has excellent denoising performance; and combined with DRCNN, it quickly reduces the computational complexity; in addition, the network architecture of the present invention adopts dimensionality reduction and dimensionality increase strategies, greatly reduces network parameters, and significantly improves the algorithm operation efficiency; it is not only suitable for denoising of known noise levels, but also for blind noise denoising of unknown noise levels, and more importantly, it is suitable for denoising of a wide range from low noise to high noise; it greatly increases the adaptability of the denoising system; at the same time, it eliminates the damage to the edge details of the denoised image, Remove boundary artifacts and improve clarity degradation; significantly improve the completeness of edge details of denoised images, and significantly improve boundary authenticity by eliminating boundary artifacts, solve the technical problem of clarity degradation and improve image clarity; perform peak signal-to-noise ratio and structural similarity combined processing; the image peak signal-to-noise ratio PNSR is significantly improved, the structural similarity index SSIM is significantly increased, the calculation efficiency is greatly improved, and the visual effect is significantly improved; the peak signal-to-noise ratio and structural similarity combined processing include: through visual characteristic analysis, using a large amount of redundant information in natural images, the areas with high contrast difference sensitivity, low spatial frequency contrast difference and brightness contrast difference exceeding the set spatial frequency contrast difference and set brightness difference are , through the smoothing of pixels in the adjacent area, the image detail texture is maintained and the visual difference is reduced; the brightness and contrast of the structural similarity undistorted image and the image to be processed are measured respectively, and the brightness measurement comparison and contrast measurement comparison are performed to obtain the brightness measurement comparison information and the contrast measurement comparison information, and the structure comparison information is analyzed; according to the brightness measurement information comparison information, the contrast measurement information comparison information and the structure comparison information, the structural similarity information of the structural similarity undistorted image and the image to be processed is obtained; the structural similarity index SSIM obtains more accurate results; the image peak signal-to-noise ratio PNSR is significantly improved, the structural similarity index SSIM is significantly increased, and the visual effect is significantly improved.

[0042] The present invention discloses a method and system for smoothing an image of a machine vision. Other advantages, objectives and features of the present invention will be reflected in part through the following description, and will also be understood by those skilled in the art through research and practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0044] Figure 1 This is a diagram showing an embodiment of an image processing smoothing method and system architecture for machine vision described in the present invention.

[0045] Figure 2 This is another embodiment diagram of the image processing smoothing method and system processing framework for machine vision described in the present invention.

[0046] Figure 3 This is a diagram showing an embodiment of an image processing smoothing method for machine vision and a system effect comparison curve according to the present invention. DETAILED DESCRIPTION

[0047] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments, so that those skilled in the art can implement it according to the instructions; as shown in the figure, the present invention provides an image processing smoothing method for machine vision, comprising:

[0048] S100, obtaining a machine vision original image, performing preliminary noise reduction processing on the machine vision original image by using a fast non-local mean algorithm and a box filter, and obtaining a preliminary noise-reduced image;

[0049] S200, the image after preliminary denoising is subjected to end-to-end low-resolution image super-resolution reconstruction through a deep residual convolutional network to obtain a super-resolution reconstructed image;

[0050] S300, performing image edge analysis and screening on the super-resolution reconstructed image to screen out edge fuzzy grayscale super-poor images;

[0051] S400, for the edge blurred grayscale out-of-parity image, the edge-preserving smoothing model is selected to remove the edge blur and restore the grayscale.

[0052] The principle and effect of the above technical solution are as follows: the present invention provides a method for image processing and smoothing of machine vision, including: obtaining the original image of machine vision, performing preliminary denoising on the original image of machine vision through a fast non-local mean algorithm and a box filter to obtain a preliminary denoised image; performing end-to-end low-resolution image super-resolution reconstruction on the preliminary denoised image through a deep residual convolutional network to obtain a super-resolution reconstructed image; performing image edge analysis and screening on the super-resolution reconstructed image to screen out edge blurred grayscale super-poor images; for edge blurred grayscale super-poor images, removing edge blur and restoring grayscale by selecting an edge-preserving smoothing model; making full use of a large amount of redundant information existing in natural images, which can better preserve the image quality. The convolutional neural network model of the large receptive field super-resolution reconstruction network fully utilizes the context information of a large number of samples to learn the implicit details and texture information in the samples, combining the advantages of the two aspects; the denoising method of the present invention has excellent denoising performance; and combined with DRCNN to quickly reduce the computational complexity; in addition, the network architecture of the present invention adopts the strategy of dimensionality reduction and dimensionality increase, which greatly reduces the network parameters and significantly improves the algorithm operation efficiency; it is not only suitable for denoising of known noise levels, but also for blind noise denoising of unknown noise levels, and more importantly, it is suitable for denoising of a wide range from low noise to high noise; it greatly increases the adaptability of the denoising system; at the same time, eliminates the denoising image The image edge details are damaged, the boundary artifacts are removed, and the clarity is improved; the edge detail integrity of the denoised image is significantly improved, and the boundary artifacts are eliminated, the boundary authenticity is significantly improved, the technical problem of clarity degradation is solved, and the image clarity is improved; the peak signal-to-noise ratio and structural similarity are combined for processing; the image peak signal-to-noise ratio PNSR is significantly improved, and the structural similarity index SSIM is significantly increased, which greatly improves the calculation efficiency and the visual effect; the peak signal-to-noise ratio and structural similarity are combined for processing, including: through visual characteristic analysis, using the large amount of redundant information in natural images, the low spatial frequency contrast difference and brightness contrast difference with high contrast difference sensitivity exceeding the set spatial frequency contrast difference and set brightness In the area with poor quality, the pixel smoothing process in the adjacent area is performed to maintain the image detail texture and reduce the visual difference; the brightness and contrast of the structural similarity undistorted image and the image to be processed are measured respectively, and the brightness measurement comparison and contrast measurement comparison are performed to obtain the brightness measurement comparison information and the contrast measurement comparison information, and the structure comparison information is analyzed; according to the brightness measurement information comparison information, the contrast measurement information comparison information and the structure comparison information, the structural similarity information of the structural similarity undistorted image and the image to be processed is obtained; the structural similarity index SSIM obtains more accurate results; the image peak signal-to-noise ratio PNSR is significantly improved, the structural similarity index SSIM is significantly increased, and the visual effect is significantly improved.

[0053] In one embodiment, S100 includes:

[0054] S101, obtaining an original machine vision image, and performing a first denoising process on the original machine vision image using a fast non-local means algorithm to obtain a first denoising processed image;

[0055] S102: Perform preliminary denoising on the first denoised image by using a box filter to obtain a preliminary denoised image.

[0056] The principle and effect of the above technical solution are: obtaining a machine vision original image, performing a first denoising process on the machine vision original image through a fast non-local mean algorithm, and obtaining a first denoising process image; according to the first denoising process image, performing a preliminary denoising process on the first denoising process image through a box filter, and obtaining a preliminary denoising process image; obtaining a machine vision original image, performing a first denoising process on the machine vision original image through a fast non-local mean algorithm, and obtaining the first denoising process image includes: for each pixel of the original image, selecting S pixels around it to calculate the average value, where S is an adaptive adjustment parameter; mapping the average value of the above S pixels to a noise-free version through a self-learning weight matrix W; normalizing the noise-free version to obtain a first denoising process image; according to the first denoising process image, performing a preliminary denoising process on the first denoising process image through a box filter, and obtaining the preliminary denoising process image includes: determining a suitable window size according to the size of the image; for each pixel on the image, setting a corresponding weight coefficient to reflect the importance of the pixel in the image. ; Input the above window size and weight coefficient into a convolution layer to obtain a preliminary denoised image; the setting method of the adaptive adjustment parameter S includes: performing brightness detection and contrast detection on each pixel of the original image and the S pixels around it and the noise-free version pixels, analyzing the brightness detection information and the contrast detection information, selecting pixels whose brightness detection information and the contrast detection information differ by more than a preset brightness difference value or a preset contrast difference value, adaptively adjusting the parameter S, and obtaining the adjusted adaptive adjustment parameter S2; mapping the average value of the adjusted adaptive adjustment parameter S2 pixels to a noise-free version through a self-learning weight matrix W; normalizing the noise-free version to obtain a second denoised image; performing preliminary denoising on the second denoised image through a box filter according to the second denoised image, and obtaining the preliminary denoised second image; constructing an image set with the preliminary denoised second image and the preliminary denoised second image to obtain a preliminary denoised image set; the original image includes: remote sensing images, medical images, astronomical images, synthetic aperture radar images, night vision images, and autonomous driving machine vision images.

[0057] In one embodiment, S200 includes:

[0058] S201, construct a deep residual convolutional network through convolution propagation stacking;

[0059] S202, inputting the image after preliminary denoising into a deep residual convolutional network;

[0060] S203, performing end-to-end low-resolution image super-resolution reconstruction through a deep residual convolutional network to obtain a super-resolution reconstructed image.

[0061] The principle and effect of the above technical solution are: constructing a deep residual convolutional network through convolutional propagation stacking; inputting the image after preliminary denoising into the deep residual convolutional network; performing end-to-end low-resolution image super-resolution reconstruction through the deep residual convolutional network to obtain a super-resolution reconstructed image; constructing a deep residual convolutional network through convolutional propagation stacking includes: setting a first convolutional layer; the first convolutional layer contains multiple 3x3 convolutional kernels for extracting high-frequency information of the image after preliminary denoising; setting a first back propagation layer for calculating the feature representation of the image after preliminary denoising; stacking the first convolutional layer and the first back propagation layer to form an end-to-end low-resolution image super-resolution reconstruction network to obtain a deep residual convolutional network.

[0062] In one embodiment, S300 includes:

[0063] S301, setting an image edge analysis screening model by identifying the image edge region;

[0064] S302, performing image edge analysis and screening on the image edge area of ​​the super-resolution reconstructed image through an image edge analysis and screening model to screen out edge fuzzy grayscale super-difference images.

[0065] The principle and effect of the above technical solution are: by identifying the edge area of ​​the image, an image edge analysis and screening model is set; the image edge analysis and screening model is used to perform image edge analysis and screening on the image edge area of ​​the super-resolution reconstructed image to screen out the edge blurred grayscale super-poor image.

[0066] In one embodiment, S400 includes:

[0067] S401, creating a selected edge-preserving smoothing model according to an edge-preserving uniform smoothing processing framework; inputting an edge-blurred grayscale super-difference image into the selected edge-preserving smoothing model;

[0068] S402, removing edge blur and restoring grayscale in the edge blur grayscale excess area by selecting an edge preserving smoothing model.

[0069] The principle and effect of the above technical solution are as follows: according to the edge-preserving uniform smoothing processing framework, a selective edge-preserving smoothing model is created; an edge-blurring grayscale out-of-tolerance image is input into the selective edge-preserving smoothing model; the edge-blurring grayscale out-of-tolerance region is subjected to edge-blurring grayscale restoration by the selective edge-preserving smoothing model; according to the edge-preserving uniform smoothing processing framework, the selective edge-preserving smoothing model is created, including: measuring the grayscale uniformity of the region by variance, setting n masks for the m×m neighborhood of any pixel on the edge-blurring grayscale out-of-tolerance image, including a k×k square, i pentagons and j hexagons; calculating the mean and variance of the n masks; without destroying the regional boundary details, the variance is sorted, and the grayscale mean of the mask area corresponding to the minimum variance is the pixel output value; eliminating noise by selective edge-preserving smoothing, finding a uniform area, and reducing sharp edge blur; finding the area with the maximum uniformity of each pixel in the surrounding image, and then replacing the original grayscale value of the pixel with the grayscale mean of the area with the maximum uniformity, and creating the selective edge-preserving smoothing model.

[0070] The present invention provides a machine vision image processing smoothing system, comprising:

[0071] The local mean box filter subsystem obtains the original machine vision image, performs preliminary noise reduction processing on the original machine vision image through the fast non-local mean algorithm and the box filter, and obtains the image after preliminary noise reduction;

[0072] Deep convolution image reconstruction subsystem: After preliminary denoising, the image is subjected to end-to-end low-resolution image super-resolution reconstruction through a deep residual convolutional network to obtain a super-resolution reconstructed image;

[0073] The edge analysis and screening subsystem performs image edge analysis and screening on the super-resolution reconstructed image to screen out images with blurred edges and grayscale out-of-parity;

[0074] The edge-preserving smoothing restoration subsystem selects an edge-preserving smoothing model to remove edge blur and restore grayscale for images with edge blur and grayscale out-of-parity.

[0075] The principle and effect of the above technical solution are as follows: the present invention provides an image processing and smoothing system for machine vision, including: a local mean box filter subsystem, which obtains the original machine vision image, performs preliminary noise reduction processing on the original machine vision image through a fast non-local mean algorithm and a box filter, and obtains a preliminary noise-reduced image; a deep convolution image reconstruction subsystem, which performs end-to-end low-resolution image super-resolution reconstruction on the preliminary noise-reduced image through a deep residual convolution network, and obtains a super-resolution reconstructed image; an edge analysis and screening subsystem, which performs image edge analysis and screening on the super-resolution reconstructed image, and screens out edge blurred grayscale super-poor images; an edge-preserving smoothing restoration subsystem, which performs edge analysis and screening on the edge blurred grayscale super-poor images by selecting an edge-preserving smoothing model. The grayscale is restored by eliminating edge blur; making full use of a large amount of redundant information in natural images, which can better maintain the details and texture of the image, and making full use of the contextual information of a large number of samples through the large receptive field super-resolution reconstruction network convolutional neural network model, learning the implicit details and texture information in the samples, combining the advantages of both aspects; making the denoising method of the present invention have excellent denoising performance; and combining DRCNN to quickly reduce the computational complexity; in addition, the network architecture of the present invention adopts the strategy of dimensionality reduction and dimensionality increase, which greatly reduces the network parameters and significantly improves the algorithm operation efficiency; it is not only suitable for denoising of known noise levels, but also suitable for blind noise denoising of unknown noise levels, and more importantly, it is suitable for denoising with a wide range of changes from low noise to high noise. ; Greatly increased the adaptability of the noise reduction system; at the same time, eliminated the damaged edge details of the noise reduction image, removed boundary artifacts and improved the clarity; significantly improved the integrity of the edge details of the noise reduction image, and through the elimination of boundary artifacts, significantly improved the boundary authenticity, solved the technical problem of clarity degradation and improved the image clarity; performed peak signal-to-noise ratio and structural similarity combined processing; the image peak signal-to-noise ratio PNSR was significantly improved, and the structural similarity index SSIM was significantly increased, which greatly improved the calculation efficiency and the visual effect; the peak signal-to-noise ratio and structural similarity combined processing included: through visual characteristic analysis, using the large amount of redundant information in natural images, the low spatial frequency contrast difference and brightness contrast difference with high contrast difference sensitivity exceeded the set For areas with spatial frequency contrast difference and set brightness difference, pixel smoothing in adjacent areas is performed to maintain image detail texture and reduce visual differences; brightness measurement and contrast measurement are respectively performed on the structural similarity undistorted image and the image to be processed, and brightness measurement comparison and contrast measurement comparison are performed to obtain brightness measurement comparison information and contrast measurement comparison information, and the structure comparison information is analyzed; structural similarity information of the structural similarity undistorted image and the image to be processed is obtained based on the brightness measurement information comparison information, contrast measurement information comparison information and structure comparison information; the structural similarity index SSIM obtains more accurate results; the image peak signal-to-noise ratio PNSR is significantly improved, the structural similarity index SSIM is significantly increased, and the visual effect is significantly improved.

[0076] In one embodiment, the local mean box filter subsystem includes:

[0077] A non-local mean fast calculation module is used to obtain a machine vision original image, and to perform a first denoising process on the machine vision original image by using a fast non-local mean algorithm to obtain a first denoising processed image;

[0078] The box filter denoising module performs preliminary denoising on the first denoising image through a box filter according to the first denoising image, so as to obtain a preliminary denoising image.

[0079] The principle and effect of the above technical solution are as follows: the local mean box filter subsystem includes: a non-local mean fast operation module, which obtains the original machine vision image, performs a first denoising process on the original machine vision image through a fast non-local mean algorithm, and obtains a first denoising process image; a box filter denoising module, which performs preliminary denoising on the first denoising process image through a box filter according to the first denoising process image, and obtains a preliminary denoising process image; the original machine vision image is obtained, and the first denoising process is performed on the original machine vision image through a fast non-local mean algorithm, and the first denoising process image is obtained, including: for each pixel of the original image, S pixels around it are selected to calculate the average value, where S is an adaptive adjustment parameter; the average value of the above S pixels is mapped to a noise-free version through a self-learning weight matrix W; the noise-free version is normalized to obtain a first denoising process image; according to the first denoising process image, the first denoising process image is preliminarily denoised through a box filter, and the preliminary denoising process image is obtained, including: determining a suitable window size according to the size of the image; for each pixel on the image, a corresponding weight is set weight coefficient to reflect the importance of the pixel in the image; input the above window size and weight coefficient into a convolution layer to obtain a preliminary denoised image; the setting method of the adaptive adjustment parameter S includes: performing brightness detection and contrast detection on each pixel of the original image and the S pixels around it and the noise-free version pixels, analyzing the brightness detection information and the contrast detection information, selecting pixels whose brightness detection information and the contrast detection information differ by more than a preset brightness difference value or a preset contrast difference value, adaptively adjusting the parameter S, and obtaining the adjusted adaptive adjustment parameter S2; mapping the average value of the adjusted adaptive adjustment parameter S2 pixels to a noise-free version through a self-learning weight matrix W; normalizing the noise-free version to obtain a second denoised image; performing preliminary denoising on the second denoised image through a box filter according to the second denoised image, and obtaining the preliminary denoised second image; constructing an image set with the preliminary denoised second image and the preliminary denoised second image to obtain a preliminary denoised image set; the original image includes: remote sensing images, medical images, astronomical images, synthetic aperture radar images, night vision images and autonomous driving machine vision images.

[0080] In one embodiment, the deep convolution image reconstruction subsystem includes:

[0081] The propagation stacking deep convolution module constructs a deep residual convolution network through convolution propagation stacking;

[0082] The residual convolutional network input module inputs the image after preliminary denoising into the deep residual convolutional network;

[0083] The end-to-end super-resolution reconstruction module performs end-to-end super-resolution reconstruction of low-resolution images through a deep residual convolutional network to obtain a super-resolution reconstructed image.

[0084] The principle and effect of the above technical solution are: a deep convolution image reconstruction subsystem, including: a propagation stacking deep convolution module, which constructs a deep residual convolution network through convolution propagation stacking; a residual convolution network input module, which inputs the image after preliminary denoising into the deep residual convolution network; an end-to-end super-resolution reconstruction module, which performs end-to-end low-resolution image super-resolution reconstruction through a deep residual convolution network to obtain a super-resolution reconstructed image; constructing a deep residual convolution network through convolution propagation stacking includes: setting a first convolution layer; the first convolution layer contains multiple 3x3 convolution kernels for extracting high-frequency information of the image after preliminary denoising; setting a first back propagation layer for calculating the feature representation of the image after preliminary denoising; stacking the first convolution layer and the first back propagation layer to form an end-to-end low-resolution image super-resolution reconstruction network to obtain a deep residual convolution network.

[0085] In one embodiment, the edge analysis screening subsystem includes:

[0086] The edge region recognition module sets the image edge analysis and screening model by identifying the image edge region;

[0087] The out-of-tolerance analysis area screening module uses the image edge analysis screening model to perform image edge analysis screening on the image edge area of ​​the super-resolution reconstructed image to screen out edge blurred grayscale out-of-tolerance images.

[0088] The principle and effect of the above technical solution are: the edge analysis and screening subsystem includes: an edge area recognition module, which sets an image edge analysis and screening model through image edge area recognition; an out-of-tolerance analysis area screening module, which performs image edge analysis and screening on the image edge area of ​​the super-resolution reconstructed image through the image edge analysis and screening model to screen out edge blurred grayscale out-of-tolerance images.

[0089] In one embodiment, the edge-preserving smooth restoration subsystem includes:

[0090] The edge uniform smoothing model module creates a selected edge preserving smoothing model according to the edge uniform smoothing processing framework; the edge fuzzy grayscale super-difference image is input into the selected edge preserving smoothing model;

[0091] The edge blur out-of-tolerance restoration module selects an edge-preserving smoothing model to remove edge blur and restore the grayscale of the edge blur out-of-tolerance grayscale area.

[0092] The principle and effect of the above technical solution are as follows: the edge-preserving smoothing restoration subsystem includes: an edge uniform smoothing model module, which creates a selected edge-preserving smoothing model according to an edge uniform smoothing processing framework; inputs an edge fuzzy grayscale super-error image into the selected edge-preserving smoothing model; an edge fuzzy super-error restoration module, which performs edge fuzzy elimination and grayscale restoration on an edge fuzzy grayscale super-error region by selecting an edge-preserving smoothing model; according to the edge uniform smoothing processing framework, creating a selected edge-preserving smoothing model includes: measuring the grayscale uniformity of the region by variance, setting n masks for the m×m neighborhood of any pixel on the edge fuzzy grayscale super-error image, including a k×k square, i pentagons and j hexagons; calculating the mean and variance of the n masks; without destroying the regional boundary details, sorting the variances, and the grayscale mean of the mask area corresponding to the minimum variance is the pixel output value; eliminating noise by selecting edge-preserving smoothing, finding a uniform area, and reducing sharp edge blur; finding the area with the maximum uniformity of each pixel in the surrounding image, and then replacing the original grayscale value of the pixel with the grayscale mean of the area with the maximum uniformity, and creating a selected edge-preserving smoothing model.

[0093] Although the embodiments of the present invention have been disclosed as above, they are not limited to the applications listed in the specification and the implementation methods. They can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily implemented. Therefore, without departing from the general concept defined by the claims and the scope of equivalents, the present invention is not limited to the specific details and the illustrations shown and described herein.

Claims

1. A machine vision image processing smoothing method, characterized in that: include: S100, obtaining a machine vision original image, performing preliminary noise reduction processing on the machine vision original image by using a fast non-local mean algorithm and a box filter, and obtaining a preliminary noise-reduced image; S200, the image after preliminary denoising is subjected to end-to-end low-resolution image super-resolution reconstruction through a deep residual convolutional network to obtain a super-resolution reconstructed image; S300, performing image edge analysis and screening on the super-resolution reconstructed image to screen out edge fuzzy grayscale super-poor images; S400, for the edge blurred grayscale out-of-parity image, the edge-preserving smoothing model is selected to remove the edge blur and restore the grayscale.

2. The image processing smoothing method for machine vision according to claim 1, characterized in that: S100 includes: S101, obtaining an original machine vision image, and performing a first denoising process on the original machine vision image using a fast non-local means algorithm to obtain a first denoising processed image; S102: Perform preliminary denoising on the first denoised image by using a box filter to obtain a preliminary denoised image.

3. The image processing smoothing method for machine vision according to claim 1, characterized in that: S200 includes: S201, construct a deep residual convolutional network through convolution propagation stacking; S202, inputting the image after preliminary denoising into a deep residual convolutional network; S203, performing end-to-end low-resolution image super-resolution reconstruction through a deep residual convolutional network to obtain a super-resolution reconstructed image.

4. The image processing smoothing method for machine vision according to claim 1, characterized in that: S300 includes: S301, setting an image edge analysis screening model by identifying the image edge region; S302, performing image edge analysis and screening on the image edge area of ​​the super-resolution reconstructed image through an image edge analysis and screening model to screen out edge fuzzy grayscale super-difference images.

5. The image processing smoothing method for machine vision according to claim 1, characterized in that: S400 includes: S401, creating a selected edge-preserving smoothing model according to an edge-preserving uniform smoothing processing framework; inputting an edge-blurred grayscale super-difference image into the selected edge-preserving smoothing model; S402, removing edge blur and restoring grayscale in the edge blur grayscale excess area by selecting an edge preserving smoothing model.

6. A machine vision image processing smoothing system, characterized in that: include: The local mean box filter subsystem obtains the original machine vision image, performs preliminary noise reduction processing on the original machine vision image through the fast non-local mean algorithm and the box filter, and obtains the image after preliminary noise reduction; Deep convolution image reconstruction subsystem: After preliminary denoising, the image is subjected to end-to-end low-resolution image super-resolution reconstruction through a deep residual convolutional network to obtain a super-resolution reconstructed image; The edge analysis and screening subsystem performs image edge analysis and screening on the super-resolution reconstructed image to screen out images with blurred edges and grayscale out-of-parity; The edge-preserving smoothing restoration subsystem selects an edge-preserving smoothing model to remove edge blur and restore grayscale for images with edge blur and grayscale out-of-parity.

7. The image processing and smoothing system for machine vision according to claim 6, characterized in that: Local mean box filter subsystem, including: A non-local mean fast calculation module is used to obtain a machine vision original image, and to perform a first denoising process on the machine vision original image by using a fast non-local mean algorithm to obtain a first denoising processed image; The box filter denoising module performs preliminary denoising on the first denoising image through a box filter according to the first denoising image, so as to obtain a preliminary denoising image.

8. The machine vision image processing and smoothing system according to claim 6, characterized in that: Deep convolution image reconstruction subsystem, including: The propagation stacking deep convolution module constructs a deep residual convolution network through convolution propagation stacking; The residual convolutional network input module inputs the image after preliminary denoising into the deep residual convolutional network; The end-to-end super-resolution reconstruction module performs end-to-end super-resolution reconstruction of low-resolution images through a deep residual convolutional network to obtain a super-resolution reconstructed image.

9. The image processing and smoothing system for machine vision according to claim 6, characterized in that: Edge analysis screening subsystem, including: The edge region recognition module sets the image edge analysis and screening model by identifying the image edge region; The out-of-tolerance analysis area screening module uses the image edge analysis screening model to perform image edge analysis screening on the image edge area of ​​the super-resolution reconstructed image to screen out edge blurred grayscale out-of-tolerance images.

10. The machine vision image processing and smoothing system according to claim 6, characterized in that: The edge-preserving smooth restoration subsystem includes: The edge uniform smoothing model module creates a selected edge preserving smoothing model according to the edge uniform smoothing processing framework; the edge fuzzy grayscale super-difference image is input into the selected edge preserving smoothing model; The edge blur out-of-tolerance restoration module selects an edge-preserving smoothing model to remove edge blur and restore the grayscale of the edge blur out-of-tolerance grayscale area.

Images