318 results about "Kernel (image processing)" patented technology

Input data comprising a video signal is processed using a combination of a known image processing method to deblur, or sharpen, the image and convolution with Pixon™ kernels for smoothing. The smoothing process utilizes a plurality of different size Pixon™ kernels which operate in parallel so that the input data are convolved with each different Pixon™ kernel simultaneously. The smoothed image is convolved with the point response function (PRF) to form data models that are compared against the input data, then the broadest Pixon™ kernel that fits the input data within a predetermined criterion are selected to form a Pixon™ map. The data are smoothed and assembled according to the Pixon™ map, then are deconvolved and output to a video display or other appropriate device, providing a clearer image with less noise.

A system and method for computer-assisted karyotyping includes a processor which receives a digitized image of metaphase chromosomes for processing in an image processing module and a classifier module. The image processing module may include a segmenting function for extracting individual chromosome images, a bend correcting function for straightening images of chromosomes that are bent or curved and a feature selection function for distinguishing between chromosome bands. The classifier module, which may be one or more trained kernel-based learning machines, receives the processed image and generates a classification of the image as normal or abnormal.

The invention provides a robust motion target detecting and tracking image processing system. The system comprises a digital signal processor taking DSP as a kernel, a CCD camera and a PC, and executes the following steps of: acquiring video data; reading acquired digital video sequences frame by frame by using the digital signal processor; performing motion target detection on the frame sequences in the last step by adopting an improved mixture Gaussian background model; performing removing noise and shadow from the detection result in the last step; tracking the detected motion target; and transmitting and displaying the processed result to a user for processing, wherein adoption of the improved mixture Gaussian background model comprises the following steps of: performing background model initialization on the input video frame to establish a background model; subtracting the current frame from a background picture; binaryzing a foreground picture; morphologically filtering to remove the noise; introducing an accelerating factor and a smooth equation to update the background model; and converting the background into an image format for display.

The invention discloses a blind camera shake deblurring method based on the L0 sparse prior, and belongs to the technical field of digital image processing. The blind camera shake deblurring method is a method for deblurring blurred images caused by camera shaking, and various space-unchanged camera shaking blurred kernels, namely the point spread functions, can be estimated. The blind camera shake deblurring method solves the problem that a current variational bayes estimation method is high in computing complexity and solves the problem that a current maximum posteriori estimation method lacks strict optimization theory supports. The blind camera shake deblurring method comprises the steps of firstly, introducing remarkable edge sparse prior based on the L0 norm, and using the iterative hard threshold compressed method to achieve recessive automatic prediction of remarkable edge characteristics, secondly, introducing camera shake blurred kernel sparse prior, and using the iterative repeated weighted least square method to achieve rapid estimation of the blurred kernels, and finally, using the image non-blind deblurring method based on super-Laplacian prior to obtain a high-quality deblurred image. The flow diagram of the blind camera shake deblurring method is shown in the figure 1.

Calculation of GR and GB color ratios in a local area uses a weighted average filter having weighting coefficients where the ratio of total sums of weighting coefficients for G and R pixels is 1:1, and weighted average filter having weighting coefficients where the ratio of total sum of weighting coefficients for G and B pixels is 1:1, respectively, on pixel lines in a horizontal direction and a vertical direction in a kernel. R and B pixel values are then calculated by interpolating the pixel value of a pixel to be processed with the G pixel value at the pixel position to be subjected to a demosaic process and the color ratio.

The invention belongs to the image processing field and relates to a multi-target tracking method and a system based on kernel function unsupervised clustering. According to the method, a binocular camera is utilized to acquire left and right sequence images at one same time, and parameters of the binocular camera are utilized for image correction; a parallax error is calculated through extracting image characteristic points and matching characteristics; the acquired parallax error is utilized to calculate the coordinate position of a target characteristic point relative to the camera, namely the coordinate of the camera, ground calibration is accomplished, ground shadow characteristic points can be filtered according to height from the characteristic point to the ground, and ground shadow interference is eliminated; according to the three-dimensional coordinate characteristic point, in combination with the kernel function, unsupervised clustering is carried out for targets with undetermined category quantity, all characteristic points of one target are gathered to form one set, one category corresponds to the position and the direction of one observation value, a present frame of the target can be acquired in combination with the position and the direction of the previous frame target, namely the prediction position value and the prediction direction value, an optimum estimation algorithm is utilized to acquire the position and the direction of the optimum target, and thereby the multi-target fast tracking effect is realized.

The present application provides an image processing method and an image processing apparatus. In the image processing method and the image processing apparatus, according to coordinate values of feature point pairs between two short exposures acquired before and after the image to be processed, so as to eliminate the influence of translating the motion coordinates of pixel points in the image tobe processed by the photographing module in the process of photographing the image to be processed, thereby improving the accuracy of the blur kernel obtained according to the motion coordinates and improving the deblurring effect of the image to be processed.

The invention discloses an image processing acceleration method and device, and an image processing model storage method and device. The acceleration method comprises the steps of obtaining an input feature map of a convolution layer in an image processing model; determining a hardware operation unit corresponding to each convolution kernel in the convolution layer from the plurality of hardware operation units; determining a convolution result between each convolution kernel and the input feature map in parallel according to a non-zero weight in the convolution kernel and the input feature map through a hardware operation unit corresponding to each convolution kernel; and generating an output feature map of the convolution layer according to a convolution result. By adopting the processing mode, the convolution operation process does not need to consider the operation of zero weight, the weight coordinate does not need to be calculated, and the data flow does not have a data dependence relationship; therefore, the network operation speed can be effectively improved, and the method is suitable for sparse CNN and non-sparse CNN at the same time.

The invention is suitable for the technical field of computer vision and image processing, and discloses a video deblurring method, device and equipment based on ambiguity, and the method comprises the steps: calculating the ambiguity of a video frame; determining a clear frame and a fuzzy frame according to the ambiguity; generating a reference frame according to the clear frame and the fuzzy frame; performing image block extraction on the fuzzy frame and the reference frame; performing weighted fusion according to the weights corresponding to the pixel points in the image block to obtain a fused image block; and recombining the fused image blocks to obtain an output image. According to the embodiment of the invention, the ambiguity of the video frame is calculated without estimating theambiguity kernel, and the clear frame and the ambiguity frame are determined according to the ambiguity, so that the calculation complexity is effectively reduced, and the calculation speed is increased; and the weight of the reference frame is considered, and weighted fusion is carried out according to the weight corresponding to the pixel points in the extracted image block, so that the definition of the finally output image is relatively high.

The invention provides an image processing system based on a convolutional neural network. The convolutional neural network at least comprises at least one convolution operation layer, at least one pooling layer and at least one full connection layer. Particularly, the image processing system comprises a development board, and a field programmable gate array FPGA and a first ARM processor are integrated on the development board. A convolution acceleration kernel is built on the FPGA. The convolution acceleration kernel is at least used for executing at least one convolution operation and at least one pooling operation and outputting a final operation result, wherein the at least one convolution operation is in one-to-one correspondence with the at least one convolution operation layer, andthe at least one pooling operation is in one-to-one correspondence with the at least one pooling layer. The first ARM processor is used for performing at least one full-connection operation on the final operation result, wherein the at least one full connection operation is in one-to-one correspondence with the at least one full connection layer. In the embodiment of the invention, the acceleration of the convolutional neural network is realized by adopting an ARM + FPGA embedded architecture.

The invention relates to a structure of an image processing unit system of a graph processor, and belongs to the field of graph processor design. The structure is composed of a front image pipeline segment unit, an image processing subclass unit, a rear image pipeline segment unit, an arbitration unit, a format conversion B unit, a compression and decompression unit and a decoding control unit. Main data including pixel, texture and bitmap can be loaded and obtained, auxiliary data including a color table and the convolution kernel is loaded and obtained, and in the data loading and obtainingprocess, functions including storage, pixel transmission, pixel mapping, pixel zooming, color table, convolution, color table transformation, column chart statistics and minimum and maximum calculation operations prescribed by the OpenGL standard are supported.

The invention discloses a two-dimensional code identification chip and a realization method therefor. The chip comprises an ARM kernel module, a storage module, a security module, an image processing module, a bus interface module and an I / O module that are all integrated on one chip and all connected with an AHB bus or an APB bus. The method comprises the steps of driving a camera to obtain a two-dimensional code image; performing blind reconstruction processing on the obtained two-dimensional code image; performing image characteristic extraction on the image subjected to the blind reconstruction processing; performing intelligent mode identification by adopting a trained SVM classifier according to the extracted image characteristics to obtain a standard type of a two-dimensional code; and performing code word extraction and decoding according to an intelligent mode identification result, and performing outputting. The chip has the advantages of low cost, high processing speed, small volume, universality, low power consumption and high reading efficiency, and can be widely applied to the technical field of two-dimensional codes.

The invention discloses a license plate positioning and recognition method based on deep learning. An improved YOLO (You Only Look Once) algorithm and an image super-resolution technology are adoptedfor optimization, and an improved YOLO convolutional neural network and a convolutional enhanced SRCNN (Super Recurrent Neural Network) convolutional neural network are trained respectively. Firstly,an improved deep learning YOLO algorithm is adopted to locate a license plate area; a correction detector is used for correcting the detection frame; according to the method, the problem that an existing license plate positioning method cannot perform correct positioning in certain specific scenes is solved, then the enhanced convolutional neural network SRCNN model is utilized to perform super-resolution technology processing on the image of the license plate area so as to obtain the picture with higher resolution and resolution rate, and then the neural network is utilized to perform opticalcharacter recognition. According to the method, when the YOLO convolutional neural network is trained, the maxout activation function is adopted to replace the activation function of the original model, so that the fitting capability is enhanced, Meanwhile, the non-maximum suppression is improved by adjusting the threshold value, so that the screening speed of the bounding box can be effectivelyincreased. When the SRCNN convolutional neural network is trained, the size of a convolution kernel and the number of convolution layers are increased, the image processing effect can be effectively improved, and therefore the method meets the requirements for real-time performance and accuracy.

The invention relates to a visible light image-based walnut maturity detection and prediction method, and belongs to the field of deep learning and image processing. The method comprises the following steps: firstly, collecting walnut samples and color images of the walnut samples in different periods, measuring fat contents of the samples, and dividing walnut maturity grades according to the fat contents and external characteristics of walnut kernels in different periods; and establishing a walnut maturity detection and prediction data set; then performing low-illumination walnut image screening and preprocessing, then inputting the images into an improved Faster RCNN network, outputting the maturity of walnuts in the images by the network, marking the maturity with a suggestion box, and evaluating the fat content of walnut kernels under the maturity at the same time. and finally, intercepting a walnut region from the original image according to the walnut suggestion box, inputting the walnut region into a walnut maturity prediction algorithm based on LSTM, and performing walnut maturity and fat content prediction after three days. The method can accurately detect the current maturity of the walnuts in the image and the maturity of the walnuts after three days and evaluate the fat content of the walnuts.

The invention relates to the technical field of image processing. The invention discloses an image processing method for resisting attacks, which comprises the following steps of: a, collecting gradient information of an image x through a local known model; b, introducing a step size amplification factor to amplify the gradient of each step in the iterative processing process, and updating the accumulative amplification gradient at the same time; c, if the cumulative amplification gradient exceeds a set threshold range, cutting noise C is obtained, and otherwise, C is 0; and d, performing projection by using a projection kernel function Wp, uniformly projecting the cutting noise C to the surrounding area of the image x, and adding the amplification gradient of the current step to obtain asample image. The method is a regional-level attack resisting technology, and provides a new thought for the research of a deep neural network. The adversarial sample image has stronger migration capability, and can better attack unknown black box models to enable the unknown black box models to generate misclassification. The technical scheme of the invention can be easily combined with many other attack methods so as to generate an adversarial sample image with stronger attack capability.

The invention discloses a fast edge-preserving filtering method for an image, and belongs to the technical field of image processing. The method comprises the following steps of: adopting a set kernel function for a filter; and accelerating a filtering process by using a FastBoxFilter to make computing time complexity be O(1), and finally obtaining a filtering result, wherein the set kernel function has four properties, namely an integral sum of 1 in a filtering template, an edge-preserving filtering property, capability of accelerating computation by using the FastBoxFilter to make the computing time complexity be O(1), and capability of adjusting parameters to achieve different filtering effects. The time complexity of the whole method is O(1), namely the time complexity is independent of the size of a filtering kernel; and moreover, effects equivalent to those of bilateral filtering can be achieved, and results show that marginal artificialities and halo phenomena are avoided.

The invention belongs to the technical field of image processing, and discloses a saturated image deblurring method and system based on a nonlinear degradation model, and the method comprises the steps: inputting a saturated blurred image or a common blurred image and a blurred kernel corresponding to the blurred image; constructing a nonlinear blurred image degradation model according to a saturated image degradation mechanism; constructing a saturated image nonlinear deconvolution framework according to the obtained degradation modeling and the maximum posterior probability framework; determining a priori item, and constructing a nonlinear energy functional model by adopting total variation priori; and solving the nonlinear energy functional through an exchange direction multiplier algorithm, namely an ADMM algorithm or a split Bregman algorithm to obtain a clear image x to be restored. According to the method, the saturated nonlinear degradation characteristic is modeled, blurring in the saturated image is removed, the ringing phenomenon caused by errors of saturated pixels is effectively restrained, and a high-quality clear image is obtained through restoration.

The invention relates to the field of defect detection, in particular to a plastic film production defect detection method based on image processing, which comprises the following steps: acquiring a plastic film grey-scale map; performing Gaussian kernel convolution on each feature extraction space in the grey-scale map to obtain each Gaussian kernel template grey-scale value; constructing a histogram in a gradient direction, and adjusting a standard deviation parameter of a Gaussian convolution function according to the histogram to obtain an adjusted gray value of each Gaussian kernel template; obtaining a gray scale change descriptor of each Gaussian kernel template according to each adjusted gray scale value, and further obtaining a gray scale change descriptor of the feature extraction space; obtaining an abnormal region in the feature extraction space according to the gray scale change descriptor of the feature extraction space; determining all defect areas according to gray level change conditions of the abnormal areas and the normal areas before and after light source enhancement; and carrying out edge detection on the defect area to obtain a defect position. The method is used for carrying out defect detection on the plastic film, and the defect detection efficiency can be improved through the method.

The invention belongs to the field of computer vision and image processing, and relates to a face super-resolution method based on a frequency decomposition multi-attention mechanism. The method comprises the following steps: performing frequency decomposition on an input low-resolution face image by utilizing the reversible property of wavelet transform and inverse transform thereof, constructing a basic module by adopting different kernel convolution aiming at the characteristics of different frequencies, integrating the features of different receptive fields adaptively, and processing the features of different frequencies by using residual attention modules including pixel, space and channel attention mechanisms, wherein low-frequency part textures adopt attention with less calculated amount, and high-frequency part adopts more residual attention modules; applying more networks to a high-frequency part while the calculated amount is kept, extracting and feeding back key points by utilizing a pre-trained face key point extraction network, enhancing contour features, and enhancing the texture features by utilizing a generated resistance network.

The invention relates to an image processing method and device, equipment and a storage medium. The method comprises the following steps: obtaining an original batch of images to be processed; according to the image processing configuration information of the convolution kernel of the neural network, obtaining a target segmentation number and a target convolution algorithm suitable for the original batch of images; equally segmenting the original batch of images according to the target segmentation number; obtaining multiple sub-batch images, and sequentially inputting the plurality of sub-batch images into a convolution kernel, so that the convolution kernel sequentially performs convolution operation on the plurality of sub-batch images by using a target convolution algorithm to obtain aplurality of sub-image processing results corresponding to the plurality of sub-batch images, and sequentially splicing the plurality of sub-image processing results to obtain an image processing result corresponding to the original batch image. According to the method and the apparatus, the sub-batch images can be subjected to rapid convolution operation in sequence by repeatedly utilizing a relatively small memory space, and the sub-batch images are spliced after operation to obtain a result equivalent to the original batch image, so that the effect of ensuring the image processing efficiency in the relatively small memory space is achieved.

The invention provides an image processing method and device and electronic equipment, and the method comprises the steps: obtaining a first weight size which is the target weight size of a convolution kernel in a convolution neural network; splitting the first weight size into second weight sizes corresponding to M * N channels, wherein the second weight size is the weight size of a convolution kernel capable of supporting convolution operation by a current hardware facility of the electronic equipment; obtaining first image data corresponding to the corresponding channel number and a secondweight size to perform convolution operation, accumulating obtained convolution results to obtain a convolution result obtained by performing convolution operation on the original image data by usingthe first weight size, the first image data being image data in the original image data, therefore, convolution operation can be performed by using the convolution kernel with any weight size in the convolution neural network model of the current equipment under the condition of not changing hardware design or slightly changing the hardware design, and the image effect output by the convolution neural network is improved.

The invention belongs to the technical field of image processing, and discloses a multi-scale fine-grained image recognition method and system based on multi-granularity attention. The method comprises the following steps: constructing an attention-based multi-granularity structure, dividing a feature extraction network into a plurality of stages, inputting images with different granularities into different stages of the feature extraction network, and carrying out feature extraction on the image to obtain an original feature map; obtaining attention weights from a channel domain and a space domain for the feature map of each stage through a multi-granularity attention module, fusing the attention weights, and then carrying out weighted fusion on the attention weights and the feature map to obtain key regions of different granularities in different stages; constructing a parallel multi-scale convolution module, grouping the feature maps, independently applying different types of convolution kernels to each group of feature maps, and performing feature extraction on the feature maps with different scales and granularities in different stages; and finally, carrying out feature fusion on the obtained feature map. The relationship between different regions can be fully mined, and low-dimensional space information and high-dimensional semantic information are fused.

The invention discloses an image processing method and device, equipment and a storage medium, and relates to the technical field of artificial intelligence, deep learning and image processing. The method comprises the following steps: inputting a mask image and a first image into a coding network; downsampling the first image by using each convolution layer of the coding network; enabling at least one convolution layer to determine a convolution kernel of each convolution window when a first image input into the convolution layer is subjected to convolution by inputting a mask image of the convolution layer; inputting a second image into a decoding network; and upsampling the second image by using each deconvolution layer of the decoding network, and outputting an image after the target area is filled. According to the embodiment of the invention, the convolution kernel of each convolution window during convolution of the first image in the same layer is determined through the mask image, so the method and the device can better adapt to the feature extraction of effective pixels in different convolution windows through employing the dynamic calculation convolution kernel, improvethe sensitivity of feature extraction, and finally enable the filling and restoration of an image missing area to be more reasonable.

The invention discloses an area nucleus segmentation counting method based on algae cell morphology, and relates to the technical field of image processing and image segmentation, and the method comprises the steps: carrying out the feature extraction and clustering of a connected region, dividing a single cell region into one class, designing a weight area kernel according to the area of the single cell region, and carrying out the matching with algae cells through the weight area kernel. Therefore, the adherent cell area is segmented, the noise area is filtered, and the purpose of counting is achieved. According to the morphological characteristics of an alga connected region, a Gaussian mixture model is used for clustering, a weight area kernel similar to an alga form is constructed, the area kernel can well filter noisy points, high precision is achieved in the aspect of segmentation of incomplete alga cells and adherent cells, meanwhile, a precise positioning mode is adopted, and the segmentation processing speed is improved.

The invention belongs to the technical field of image processing, which discloses an image interpolator combining a plane interpolation and a spherical surface interpolation and a method, wherein the image interpolator consists of a plane interpolation module and a spherical surface interpolation module; an input end of the plane interpolation module and the input end of the spherical surface interpolation module are respectively connected with an output end of an image local arbiter; and high resolution ratio image signal output ends of the plane interpolation module and the spherical surface interpolation module after image interpolation processing are connected with an image display or a storage. The method comprises the following steps: a digital image is equivalent to a point set in a space coordinate system; the space relation of four points around a point to be interpolated is built; the four points are on the same plane or on the same spherical surface; and the pixel value of the point to be interpolated is determined. The defects that the same primary function or interpolation kernel is adopted for approaching an original image in the traditional image interpolation method are avoided through the image interpolator and the method; and the invention has the advantages of the best image effects, clear images, high resolution ratio and zero distortion after the interpolation processing.

The invention discloses an X-ray breast mass image automatic classification method. According to the invention, an automatic classification network for the X-ray breast mass image is designed from theperspective of image processing; according to the network, firstly, two computing paths are used for carrying out convolution and downsampling operation on an X-ray breast mass image by using convolution kernels of different sizes, convolution feature maps of different scale types are extracted, the feature maps input by the two computing paths are superposed and fused, and feature information obtained after double computing paths are fused is obtained. Feature extraction is carried out on the fusion features by using a full convolutional network, and finally the extracted features are sent to a Softmax classification layer to classify the features, and a breast mass image classification result is obtained. A model is trained by using a membership-based objective function suitable for X-ray breast mass image classification, and a new objective function enhances the generalization ability of the model by increasing the membership degree of a breast mass sample and a category to which the breast mass sample belongs and reducing the membership degree of the breast mass sample and a non-category to which the breast mass sample belongs, so that the classification accuracy is improved.

The invention provides an image processing method and device, electronic equipment and a storage medium. The method comprises the steps of inputting a to-be-processed image into a pre-trained convolutional network model; processing the to-be-processed image through the convolutional network model; and outputting a processing result of the to-be-processed image. The convolutional network model comprises a convolutional network module with linear phase constraint. The convolutional network module comprises a deep convolutional network layer and a linear phase point-by-point convolutional networklayer. The deep convolution network layer adopts a convolution kernel with the size of 3*3. The linear phase point-by-point convolution network layer adopts a linear phase point-by-point convolutionkernel with the size of 1*1, and the weight of the linear phase point-by-point convolution kernel is symmetric or antisymmetric in the depth direction. According to the invention, the image is processed based on the convolutional network model with the channel number reduction function, so that the parameter quantity in the image processing process can be effectively reduced, and the image processing complexity and the calculation amount of image processing equipment are reduced.

The invention belongs to the field of image processing, and particularly relates to a learning-based nonlinear dynamic system deblurring method. The method comprises the steps that firstly, for an image to be deblurred, the kernel estimation energy is controlled by a learnable non-linear dynamic system; secondly, after continuous iteration of a latent image and a fuzzy kernel, great fuzzy kernel estimation is acquired; and finally, the problem of blind deblurring is transformed into the problem of non-blind deblurring, and various existing non-blind deblurring methods can be used for solving.The method has the advantages that 1 a new principle for deblurring is provided, and the learnable dynamic system instead of manually set regularization is used to control kernel estimation; 2 a new structure is designed to learn components in the dynamic system, and a suitable and flexible deblurring system is acquired through the structure; and 3 the method involves a residual network proposed recently, which brings new ideas to image processing and deep learning.

The embodiment of the invention provides an image processing method and device, electronic equipment and a computer readable storage medium. The method comprises the steps: obtaining a to-be-processedimage and a zoom ratio value of the to-be-processed image; performing feature extraction on the to-be-processed image to obtain a first resolution feature map and a corresponding size; performing convolution kernel prediction processing according to the size of the first resolution feature map and the zoom ratio value to obtain a predicted convolution kernel, and performing convolution processingon the first resolution feature map by using the predicted convolution kernel to obtain a second resolution feature map; and performing feature extraction processing on the second resolution featuremap to obtain a second resolution image. According to the invention, a resolution reconstruction task with any zoom magnification can be realized, the artifact problem caused by directly outputting the image after the resolution is adjusted is avoided, and the quality of the image after the image resolution is adjusted is improved.

The invention relates to the field of image processing algorithms. According to the technical scheme, the famous high-quality tea picking point position information obtaining method based on machine vision comprises the following steps: 1, obtaining a tea image from a tea garden, and performing Gaussian filtering noise removal on the tea image through a 3 * 3 convolution kernel; 2, setting a respective ROI for each tender shoot obtained through the image; 3) converting the ROI in the RGB color space into an HSV color space, and extracting the features of the tender shoots and the growing point branches of the tender shoots; 4) carrying out secondary binarization segmentation on the extracted tender shoot and branch areas by using an Otsu algorithm; 5) refining the binarized picture in the previous step by adopting an improved Zhang refining algorithm, and extracting a skeleton of the binarized picture; 6) using a Shii-Tomasi algorithm to search bifurcation points of the tender shoots and the branches as feature angular points for detecting the refined skeleton; 7) fitting the lowest point and the angular point of the contour into a linear segment. According to the method, the precision and efficiency of tea tender shoot picking point positioning can be improved.