41results about How to "Preserve Texture Details" patented technology

The invention discloses an image processing method and device and electronic equipment. The image processing method comprises that edge-maintaining filtering is carried out on a first image based on an original image to obtain a filtering image; and a second image based on the original image is fused with a third image based on the filtering image to obtain a fusion image. Thus, texture details ofthe original image can be maintained to certain extent while noise is eliminated and the edges are maintained, image distortion due to the fact that a flat area is too smooth can be avoided, the fusion image approaches the reality more, and the image quality can be improved effectively.

A printing defect detection method based on a flexible template registration comprising the following steps: adopting a rigid registration for initial positioning; using a training sample subjected to rigid initial registration as well as overall similarity measure and coordinate difference degree of all pixels of a standard image as an optimization goal to acquire a flexible registration image from the standard image to the training sample; obtaining a high-value image and a low-value image by studying a training sample set subjected to flexible registration; and finally determining whether the grayscale and / or the color of all images to be detected exceed(s) the value range of high and low template or not so as to obtain a defect image. The method can ensure the image detection accuracy while maintaining tolerance on destabilizing factors such as deformation, so that not only can be the misinformation probability of machine vision inspection reduced, but also the risk of leakage waste can be reduced, which is greatly conducive to improving the detection accuracy.

The invention provides an SAR-fused visible light remote sensing image defogging method, which specifically comprises the following steps of: step 1, selecting SAR images and optical remote sensing images in the same region in the same time period so as to construct an optical-SAR remote sensing image defogging data set, and dividing the data set into a training set and a verification set; step 2,constructing a conditional generative adversarial convolutional neural network as a defogging model, wherein the model is composed of a generative network and a discrimination network; step 3, jointly training the defogging model; and step 4, defogging a visible light remote sensing image by using the trained defogging model. According to the SAR-fused visible light remote sensing image defoggingmethod, the mapping relation between a foggy image and a clear image is directly learned through using the conditional generative adversarial network, and end-to-end defogging can be achieved; SAR information is fused, and defogging visual enhancement is achieved; and a cascaded residual expansion convolution block structure is adopted, so that supervised learning can be performed on the defogging model, and efficient defogging can be realized.

The invention discloses a texture rendering method and system for real-time three-dimensional human body reconstruction, a chip, equipment and a medium. The method comprises the steps of obtaining a current human body model and a depth image of a shooting object; selecting a current human body model as a standard model, reprojecting the vertex of the standard model to the depth image, extracting color information and image coordinates corresponding to the vertex, wherien the color information is a color initial value, and the image coordinates are converted into texture coordinates; calculating a weighted sum of the subsequent color information of the vertex of the human body model and the color initial value to serve as a new color of the vertex of the standard model; calculating sub-texture maps and sub-masks of the current human body model, and combining the sub-texture maps and the sub-masks into a complete texture map and mask; and performing rendering according to the texture mapand the texture coordinates. According to the method, generation and optimization of required textures can be rapidly completed based on the GPU, a high-quality texture atlas is obtained, and color cracks caused by illumination changes are eliminated. A human body model generated in a multi-camera system can be rendered, and a good visual reality sense is achieved.

The invention relates to an angular super-resolution reconstruction method on the basis of perception loss for images of light fields, and belongs to the field of light field imaging. The angular super-resolution reconstruction method includes that mean square errors of image high-dimensional features extracted by pre-training models are used as loss functions, a network model which comprises fourresidual blocks is built, nonlinear mapping relations between observation images and target viewing angle images are learned, and accordingly new viewing angle images can be reconstructed. The angular super-resolution reconstruction method has the advantages that the perception loss for expressing the high-dimensional features is introduced, accordingly, texture details of the super-resolution reconstructed new viewing angle images can be effectively kept, and excellent visual effects can be realized by the angular super-resolution reconstruction method.

The invention discloses an image reconstruction method based on group sparsity coefficient estimation. The method belongs to the technical field of digital image processing and is an image reconstruction method based on similar block set sparsity coefficient estimation, wherein similar image blocks are searched by an Euclidean distance at first; partial and non-partial sparse representation is carried out to a similar image block set, so that the sparser and more accurate coefficient is obtained; a reconstruction model is solved further by a Bergman iteration algorithm; and the sparsity coefficient is estimated according to a linear minimum mean square error rule, so that the accurate estimation of the small coefficient containing image texture detailed information is ensured. The method disclosed by the invention has the advantages that the linear minimum mean square error estimation is carried out to the similar image block set sparsity representation coefficient, so that obvious effects are obtained in repair, deblurring and other aspects; a reconstructed image will also have the more abundant detailed information; overall visual effects become clearer; and the method can be used in the repair and the deblurring of the optical images.

The invention discloses a suspected lung nodule image enhancement directional scale filtering method, which comprises the following steps: inputting an original image, and calculating a peak value image and a valley value image of the original X-ray chest radiography respectively by utilizing a top-hat transform operator and a bottom-hat transform operator in gray morphology; carrying out directional enhancement on a lung similar-round focus image in the X-ray chest peak value image by utilizing a directional scale Laplace Gauss function as a matched filter, introducing a visual correction factor in image enhancement operation, and carrying out visual correction on a convolution image of a directional scale Laplace Gauss operator; and adding the original X-ray chest image in the image obtained from the step 2, and then subtracting the valley value image obtained from the step 1 to obtain an output image. The beneficiary effect of the invention is that the grey contrast between the lung similar-round structural focus in the X-ray chest image and the periphery background tissue of the lung similar-round structural focus is enhanced obviously, the texture details in the X-ray chest are maintained, and the visual effect of the X-ray chest is improved preferably.

The invention relates to an NSST domain flotation froth image enhancement and denoising method based on a quantum harmony search fuzzy set. The NSST domain flotation froth image enhancement and denoising method comprises the steps: carrying out NSST decomposition on a flotation froth image, and obtaining a low-frequency sub-band image and multi-scale high-frequency sub-bands; performing quantum harmony search fuzzy set enhancement on the low-frequency sub-band image; secondly, for the multi-scale high-frequency sub-bands, removing a noise coefficient by combining an improved BayesShrink thresholding and scale correlation, and enhancing an edge coefficient and a texture coefficient through a nonlinear gain function; and finally, performing NSST reconstruction on coefficients of the processed low-frequency sub-band and each high-frequency sub-band to obtain an enhanced de-noised image. According to the NSST domain flotation froth image enhancement and denoising method, the brightness, the contrast and the definition of the froth image can be improved, and the froth edge is obviously enhanced while noise is effectively inhibited, and more texture details are reserved, and subsequent processing such as froth segmentation and edge detection is facilitated.

The invention discloses an image denoising method based on a fractional order partial differential equation. According to the method, based on the non-local property of a fractional order derivative, the interference of noises can be alleviated during detection of edges, the image denoising method based on the fractional order partial differential equation is obtained with the combination of the partial differential equation, denoising is realized, and texture details of an original image can be reserved as many as possible; in the solving process, a method of fast Fourier transform is employed, expanded operation of complicated fractional order derivatives is avoided, and the solving speed is accelerated; and the fractional derivative is individually set for a variable of a spread function, good denoising effect can be achieved for different image variation differential orders, the convergence speed is fast, and the required iteration frequency is low.

The invention discloses a polarimetric SAR classification method on the basis of NSCT and discriminative dictionary learning and mainly solves the problems of low classification accuracy and low classification speed of an existing polarimetric SAR image classification method. The polarimetric SAR classification method comprises the following implementing steps: 1, acquiring a coherence matrix of a polarimetric SAR image to be classified and carrying out Lee filtering on the coherence matrix to obtain the de-noised coherence matrix; 2, carrying out Cloude decomposition on the de-noised coherence matrix and using three non-negative feature values of decomposition values and a scattering angle as classification features; 3, carrying out three-layer NSCT on the classification features and using a transformed low-frequency coefficient as a transform domain classification feature; 4, using the transform domain classification feature and combining a discriminative dictionary learning model to train a dictionary and a classifier; 5, using the dictionary and the classifier, which are obtained by training, to classify a test sample so as to obtain a classification result. The polarimetric SAR classification method improves classification accuracy and increases a classification speed and is suitable for image processing.

The invention discloses a self-adaptive target detection method based on an SPCNN, and belongs to the technical field of computer vision. The implementation method comprises the following steps: calculating an image static attribute parameter; deducing a theoretical formula according to the Stevens law, and calculating a threshold attenuation time constant [alpha]e, so that the threshold attenuation time constant [alpha]e can be adaptively set according to the overall gray feature of the target image; based on an adaptive side inhibition mechanism, improving an inhibition coefficient calculation model by using a hyperbolic tangent function, and calculating a link weight matrix of each pixel point by using the inhibition coefficient calculation model; inputting the image into an SPCNN with complete self-adaptive setting of parameters, continuously iterating and generating a binarization segmentation result, and extracting candidate targets. Based on a fast connection mechanism in neuron synchronization, in combination with a grayscale image criterion, automatic output of an optimal segmentation result is realized by calculating the similarity of adjacent iteration segmentation results and searching a similarity maximum value, and meanwhile, iteration is automatically controlled, so that the efficiency and intelligence of a target detection method are improved.

The invention provides an adaptive partitioning method for pulse array coding. The method comprises the following steps: time-space pulse information is acquired by a time-space signal sensor to forma pulse array; the pulse array is partitioned into multiple coding tree cubes; each coding tree cube is partitioned into coding cubes through multi-level partitioning until being partitioned to the maximum depth; the performance of the coding cube structure before partitioning each time is compared with the performance of the coding cube structure after partitioning to decide whether or not to conduct code cube partitioning; and partitioning structures of the coding tree cubes and the coding cubes are determined, and partitioning results are output. According to the method, an adaptive partitioning mode for a time-space pulse array signal with both time and space is proposed according to a partitioning thought for a coding structure in traditional video coding; and by means of partitioningin the space domain and the time domain, the operated area range is provided for the follow-up compression process.

The invention provides an image rain removing method and a device, comprising: an image to be detected with raindrops is separated into a high-frequency component image and a low-frequency component image; the high-frequency component image is input to the trained convolution neural network with residual structure, and the rain-free image is outputted; the rain-free image is synthesized with the image to be detected, and the rain-free image is obtained; the rain-free image is inputted to the trained convolution neural network with discriminant structure, and the category of the rain-free imageand the raindrop-free image corresponding to the image to be detected are outputted; the category is inputted to the convolution neural network with a residual structure, and the parameters of the convolution neural network with residual structure are updated to obtain the final rain-free image. The final rain-free image obtained by the invention retains the texture details of the rain-free areain the image to be detected, so that the final rain-free image is close to the rain-free image corresponding to the image to be detected, and the important factors in the image to be detected are wellretained.

The invention provides a large-scale liftoff distance satellite-borne SAR image mosaicking method, and relates to the field of satellite SAR (synthetic aperture radar) image mosaicking. The method comprises the following steps of: carrying out necessary preprocessing on an original image; extracting geographic location information of the preprocessed image; carrying out correlated correction on the image by taking the geographic location information as a priori knowledge so as to obtain mosaicking error information; compensating an image mosaicking error by utilizing the mosaicking error information; after the mosaicking error compensation, carrying out uniform color processing on the input SAR image in a search updating manner by utilizing a Wallis filter; and finally generate a mosaickedimage. According to the method, a cross-correlation method or zero crossing point-based binary matching statistical method is adapted according to different terrains, so that selectable schemes are provided for the determination of matching offsets; and a search updating strategy is provided, so that a uniform color processing effect at multiple overlapping areas in multiple directions is ensured, and texture details in the images can be well protected.

The invention discloses an image denoising method. The method includes the steps of initializing parameters and image inputs, performing discrete Fourier transformation on initialized images, solving a fractional diffusion-wave equation model of a discrete space based on iterative computations, performing discrete Fourier transformation on iterative results to obtain a final result. By the adoption of the method, real edge structures and important texture details of the images can be well maintained, and the method is superior to similar methods evaluated both from objective performance indexes and subjective visual perception.

The invention provides an image enhancement method and device and a computer readable storage medium, and belongs to the technical field of image processing. The image enhancement method comprises the following steps: acquiring multiple groups of training data, wherein each group of training data comprises a first image and a second image; constructing an image enhancement model comprising a first-stage generative adversarial network and a second-stage generative adversarial network, training the image enhancement model by using multiple groups of training data, and training the first-stage generative adversarial network by using an enhanced low-frequency image generated based on the low-frequency features of the first image by taking the second image as a target image; taking the second image as a target image, generating an enhanced image based on the fused image of the first image, and training the second-stage generative adversarial network; the fused image is obtained by fusing the first image and the enhanced low-frequency image; and inputting a third image to be enhanced into the trained image enhancement model, and outputting a fourth image after image enhancement. The image quality of the image can be improved.

The invention provides an image processing model training method, a document image processing method and equipment. The method comprises the following steps: respectively adding noise into a plurality of noise-free white-background black character images to obtain a plurality of noisy white-background black character images; respectively fusing the noisy white-background black character image with a plurality of background images to obtain a plurality of sample document images, the plurality of background images being a plurality of different scene images; performing color inversion processing on the noise-free white-background black character image to obtain a noise-free black-background white character image; the sample document images and the noise-free black-matrix white character images being in one-to-one correspondence, obtaining a training sample set, the training sample set comprising multiple sets of sample images, and each set of sample images comprising one sample document image and the corresponding noise-free black-matrix white character image; and carrying out model training according to the training sample set to obtain an image processing model so as to carry out de-noising and blackening and whitening processing on an input document image. The image processing model trained by the invention can better retain the text content and can process a plurality of document images at the same time.

The invention discloses a morphological attribute filtering multimode fusion imaging method, which comprises the steps of performing morphological attribute filtering operation on a to-be-fused infrared image, solving an adaptive segmentation threshold, and performing binarization to obtain an infrared image weight map; carrying out edge preserving filtering on the infrared image weight map; calculating according to the infrared image weight graph to obtain a visible light image weight graph, and respectively constructing an image pyramid for the visible light image to be fused, the infrared image to be fused, the infrared image weight graph and the visible light image weight graph; fusing the visible light image, the infrared image and the weighted image pyramid to be fused to obtain fusion, and reconstructing to obtain a final fusion result. According to the morphological attribute filtering multi-mode fusion imaging method, quick and stable multi-mode fusion imaging can be performedby utilizing an image processing means, a fusion imaging result can effectively reserve a salient target in an infrared image and edge and texture details in a visible image, Meanwhile, the morphological attribute filtering multi-mode fusion imaging method has the advantages of high calculation efficiency and good universality.

The invention discloses a method, system and device for detecting an SAR image based on ROF model semi-implicit denoising. The method is characterized in that the method comprises the steps: carryingout the logarithmic transformation of two time phase images with noise, and removing the noise of the images with noises through semi-implicit difference scheme via an ROF model after logarithmic transformation; carrying out the difference calculation of the two time phase images with noise, and obtaining a difference image; carrying out the clustering of the difference image, and obtaining a change detection result graph. The method achieves the more accurate and complete obtaining of the change information of a remote sensing image.

The invention relates to a threshold segmentation method and system based on a region variance weight, and the method comprises the steps: obtaining a first threshold Th1 through employing an improved Otsu threshold segmentation algorithm, segmenting an image into a background region and a target region based on the first threshold Th1, and calculating the variance of the two segmented regions; solving a second threshold Th2 by using a maximum entropy threshold segmentation algorithm, segmenting the image into a background region and a target region based on the second threshold Th2, and calculating the variance of the two segmented regions; and then, a weight coefficient is introduced to balance the size of the proportion of the four variances, the size of two threshold values is adaptively adjusted through the weight coefficient, so that an accurate segmentation threshold value Th is obtained, and the image is segmented into a background region and a target region based on the segmentation threshold value Th. The method and the system are beneficial to accurately segmenting the image into the target area and the background area without losing image texture details.

The invention relates to a regional aware image de-noising method based on machine learning, comprising the following steps: 1. using a noise standard deviation [sigma] and a standard deviation rj*[sigma] reduced by k-type reduction ratio as noise de-noising parameters to obtain different de-noising result sets; 2, combining the [sigma] with the de-noising results using rj*[sigma] to obtain the preference of the optimal reduction ratio r<~> and an image block using the two de-noising parameters using [sigma] and r<~>*[sigma]; 3, performing feature extraction on the noise image and the de-noising results using the two de-noising parameters; 4, using an obtained preference feature set as a feature set of a machine learning algorithm to learn the de-noising parameter preference model of the image block; 5, using the de-noising parameter preference model to predict the noise image in the test set to obtain the predicted preference probability value of each image block; and 6, performing threshold processing and combining the de-noising results of the two de-noising parameters to obtain a final de-noising result. The method can improve the performance.

The invention discloses a fabric image recoloring method, wherein the method comprises the steps: step 1, segmenting a foreground image and a background image of a fabric pattern, and segmenting the foreground and the background of the fabric pattern by using a color feature coding algorithm, an edge structure image extraction algorithm and a multi-region fuzzy competition image segmentation model; step 2, coloring the background image of the fabric pattern, and coloring the background of the fabric pattern by adopting a coloring algorithm based on a grayscale image; and step 3, coloring the foreground image of the fabric pattern, for the foreground of the fabric pattern, reconstructing the foreground of the fabric by adopting a color offset field model based on intrinsic image decomposition, obtaining a color center vector and a color offset field of the foreground pattern of the fabric, and for the reconstructed foreground image, replacing the original color center vector of the foreground pattern with a given new color center vector, and reconstructing a new fabric pattern by using a color offset field. The invention further comprises a fabric image recoloring system.

The invention provides an adaptive partitioning method for pulse array coding. The method comprises the following steps: time-space pulse information is acquired by a time-space signal sensor to forma pulse array; the pulse array is partitioned into multiple coding tree cubes; each coding tree cube is partitioned into coding cubes through multi-level partitioning until being partitioned to the maximum depth; the performance of the coding cube structure before partitioning each time is compared with the performance of the coding cube structure after partitioning to decide whether or not to conduct code cube partitioning; and partitioning structures of the coding tree cubes and the coding cubes are determined, and partitioning results are output. According to the method, an adaptive partitioning mode for a time-space pulse array signal with both time and space is proposed according to a partitioning thought for a coding structure in traditional video coding; and by means of partitioningin the space domain and the time domain, the operated area range is provided for the follow-up compression process.

The invention belongs to the field of video processing, in particular relates to video enhancement technology, and specifically provides a video denoising method based on prior information and a convolutional neural network. The present invention denoises the noise video based on the convolutional neural network, and constructs a denoising neural network composed of two parts connected, wherein the first part is a 4-layer 1×1 convolution kernel connected in sequence, and each convolution kernel The ReLU activation function is connected; the second part is a 15-layer 3×3 Octave convolution kernel connected in sequence, and the first to 14th layer convolution kernels are connected with batch normalization and ReLU activation functions; at the same time, fully in the training set construction and waiting time During the preprocessing process of noise video data, the front and back frame information is fully utilized. To sum up, compared with the traditional method, the present invention does not need to manually adjust parameters, has a good denoising effect, can well maintain texture details in the video, is easy to use, runs fast, and has high robustness.

The invention relates to the field of image processing, in particular to a motor train unit sanding pipe joint disconnection fault detection method based on image processing, solves the problems of missing detection and wrong detection caused by the fact that whether a sanding pipe joint is disconnected or not is checked in an existing manual image checking mode, and relates to the field of imageprocessing. The method comprises the steps: taking the features of sanding pipe component images as training features, and acquiring a trained classifier; processing the to-be-detected image containing the motor train unit sanding pipe to obtain a noise-free image containing the motor train unit sanding pipe; matching the noiseless image with a sanding pipe template image stored in a template image library, and extracting a sanding pipe joint image from the noiseless image according to the successfully matched sanding pipe template image; extracting features of the sanding pipe joint image asto-be-detected features; inputting the to-be-detected features into the trained classifier, and outputting the category of the sanding pipe joint image. The device is used for identifying whether thesanding pipe joint is disconnected.

The invention relates to a self-adaptive non-local mean filtering method for salt and pepper noise, and belongs to the technical field of digital image processing. The invention provides a simple and effective method for removing salt and pepper noise. Firstly, a sliding window is used to identify noise points, and local filtering is used to carry out preliminary denoising. Secondly, non-local mean filtering with adaptive parameters is provided for secondary denoising; a smoothing parameter is designed as a piecewise function according to the intensity level of the salt and pepper noise. Experimental results on a public data set show that the novel filter balances the relationship between the denoising effect and the consumed time. Moreover, the novel filter can effectively restore the pixels of a polluted image and retain the texture details of the image.