88 results about "Gaussian convolution" patented technology

The invention relates to a hybrid image processing method of extracting the centric line of the structured-light stripe based on the ROI area, belonging to the field of machine vision technique. By combining the image threshold processing, communicated components extraction, shape factor constraint and expansion method, the invention automatically divides the structured light area as the ROI area of light stripe extraction, using the extraction method of fringe center based on Hessian matrix to realize the extraction of sub-pixel image space of the centric line of the structured-light stripe of the ROI area. The method has the advantages of high precision, good robustness, high automation degree and other advantages, avoiding the large formwork Gaussian convolution calculation of the pixel the non-light stripe area, thereby reducing the redundant calculation of the extraction of the structured light stripe and realizing the high speed and precision extraction of the centric line of the light stripe, laying a foundation for the real-time application of the structured light 3D vision measurement.

The invention provides a Retinex image enhancement method based on a Laplacian pyramid. The method comprises the following three processes: firstly, carrying out multi-scale convolution on an originalimage, carrying out convolution by adopting three 3 * 3 Gaussian convolution kernels, and obtaining R, G and B channel feature maps with different scales; secondly, estimating an incident component by adopting a Retinex algorithm and enhancing colors; enhancing details of reflective components; and finally, carrying out linear weighted fusion on the color enhanced image and the detail enhanced image to obtain an enhanced image. According to the method, more details of the image are acquired by adopting multi-scale convolution, image enhancement is carried out by utilizing an MSRCR algorithm,and Gamma correction is carried out on an incident component to realize color enhancement. Bilateral filtering denoising is carried out on the reflection component to reduce noise interference, and image detail information is extracted through a Laplacian pyramid; and linear weighted fusion is carried out on the color enhanced image and the detail enhanced image to realize image enhancement.

The invention relates to a level-set retinal vessel image segmentation method with fusion shape prior, comprising: (1) utilizing morphological operators and Gaussian convolution to enhance the retinal vessel image; (2) adopting the anisotropy characteristic of Hessian matrix and improving The vascular response function of the retinal vascular image is roughly segmented and used as shape constraints and initialization information; (3) using the shape prior and retinal image data information to construct a model that includes local area energy fitting items, shape constraint items, and level set function regularity maintenance A level set model for retinal vessel segmentation with length penalty term, weighted area constraint term. The segmentation result of the present invention has high accuracy, can replace manual segmentation, can play an important auxiliary role in the diagnosis and treatment of clinically relevant ophthalmic diseases, and has strong clinical application value.

The invention belongs to the technical field of computer vision and relates to a 2D to 3D conversion method, in particular to a single image-based 2D to 3D conversion method. The method comprises the following four steps: subjecting an image to be processed to grayscale preprocessing; calculating a two-dimensional Gaussian convolution mask according to a two-dimensional Gaussian filtration formula; performing the Gaussian smoothing filtration of the grayscale image by using the obtained two-dimensional gauss convolution mask to reduce the noises in the image; and performing the Laplace filtration of the image subjected to the two-dimensional Gaussian filtration. The method has the advantages of processing by starting from the image, avoiding the complex input of multiple images or multi-view stereo images, having high applicability, not requiring camera parameters, avoiding complex calibration, correction and other operation, ensuring automatic generation in a whole process, requiring a small amount of calculation, and having convenient application on occasions when high speed is required.

The invention relates to a method and a device for realizing Gaussian blur, belonging to the technical field of image processing. The prior art has the defects that the calculated amount for realizing Gaussian blur is large and the processing speed is slow. The method of the invention comprises that horizontal one-dimensional Gaussian convolutions and vertical one-dimensional Gaussian convolutions are sequentially conducted to images to be processed. The invention additionally discloses a device for realizing the method. The device comprises an external input receiving unit, a convolution kernel generation unit, a horizontal convolution unit and a vertical convolution unit, wherein the external input receiving unit is used to determine the width of a convolution kernel according to the processing width of the received images, the convolution kernel generation unit is used to determine the specific values of the convolution kernel according to the width of the convolution kernel and the Gaussian function, the horizontal convolution unit is used to conduct horizontal one-dimensional Gaussian convolutions to each line of the images to be processed, and the vertical convolution unit is used to vertical one-dimensional Gaussian convolutions to the results of the horizontal one-dimensional Gaussian convolutions. By adopting the method and the device of the invention, the calculated amount for realizing Gaussian blur is reduced and the processing speed is improved.

The invention relates to processing of a medical image and in particular relates to a method for enhancing the liver blood vessel and simultaneously dividing the liver from the blood vessel in a CTA (computed tomography imaging) image. The method comprises the following steps of: preprocessing an image by means of gaussian convolution; computing the anisotropic characteristic value of each point of the image, and further conforming the anisotropic oval neighbourhood of each point; computing a grey level histogram in each neighbourhood, initiating a liver region, and computing a grey level histogram in the liver region; computing the wasserstein distance between the grey level histogram in each point neighbourhood and the grey level histogram in the liver region; enhancing the blood vessel according to the wasserstein distance and the anisotropic characteristic of the neighbourhood; and dividing the liver from the blood vessel. According to the method provided by the invention, the negative effects caused by the low contrast ratio, the noise, the fuzzy boundary and the like can be overcome, and the blood vessel identifying and dividing accuracy rate can be greatly improved, thus the anatomical structure information of the liver blood vessel can be exactly obtained.

The invention discloses a color-information-based scale invariant feature point describing and matching method which comprises the following steps: firstly, carrying out Gaussian convolution on R, G and B chrominance components in digital color images respectively; according to the coordinates, direction, scale and other information of feature points, determining the position and structure of feature point descriptors; calculating the average value of the R, G and B chrominance components of subregions in the concentric circle structure of the descriptor, and taking each average value as one-dimensional element of the feature vector to construct a feature vector; according to the distance between the subregion and the feature points, multiplying all-dimensional feature vector elements by Gaussian weight; respectively carrying out normalization treatment on the feature vector element belonging to the same one chrominance component; sequentially calculating the feature vectors of all feature points, and constructing feature vector space of the images; and finally calculating the distance between every two feature vectors and matching corresponding feature points in the feature vector space of the two images.

The invention discloses a Harris corner detecting parallel software system based on the design idea of a Graphics Processing Unit (GPU). The time-consuming image Gaussian convolution smoothing filtering part in calculation is improved to a Single Instruction Multiple Thread (SIMT) mode through a plurality of threads, and the whole process of image corner detecting is finished on a Compute Unified Device Architecture (CUDA) through a shared memory, a constant memory and a lock page internal storage mechanism in the GPU. Corners detected by the software system are distributed evenly, and good effects in the aspects of corner extracting and precise positioning are achieved. Compared with a serial algorithm based on a CPU, a Harris corner detecting parallel algorithm based on the GPU can obtain a speed-up ratio reaching up to 60 times. The executing efficiency of the Harris corner detecting parallel algorithm based on the GPU is improved obviously, and the good real-time processing capacity is shown in the aspect of large-scale data processing.

The invention discloses an intelligent face skin aging identification and evaluation method. The method comprises the steps of image collection and preprocessing, namely collecting a face image used for assessing the aging degree of the face skin, converting the image into a grayscale image from an RGB image, carrying out Gaussian convolution on the grayscale image, and carrying out Hessian matrixalgorithm calculation to obtain a binary image; wrinkle detection and screening, namely judging whether the region with the value of 1 in the binary image is a suspected wrinkle region, judging whether the suspected wrinkle region is a real wrinkle or not according to the minimum bounding rectangle of the region and the inclination angles of the long edge, the short edge and the long edge of theregion, and processing to obtain a new wrinkle binary image; extraction of wrinkle features. Wherein the region with the value of 1 in the wrinkle binary image is a real wrinkle region, and extractinghuman face wrinkle characteristics including the number of wrinkles, the maximum communicating length of the wrinkles, the maximum width, the wrinkle color depth degree and the minimum circumscribedrectangular area of the wrinkles; evaluation of human face skin aging degree. And the human face skin aging degree and the human face skin visual aging degree are obtained through weighted comprehensive calculation of the human face wrinkle characteristics. According to the method, the human face skin aging degree is comprehensively judged by rapidly and accurately detecting the human face wrinkleindex, and the quantitative index is provided.

The invention discloses an image FFT symbol information based unmanned aerial vehicle autonomous landing target detection method. The method comprises the steps of (1) converting RGB color space of an obtained image Inum into LAB color space; (2) performing FFT to obtain an image Inum FFT processing result Inum_FFT; (3) performing symbol information extraction on the transformed image, that is, performing symbolic operation on the image Inum_FFT to obtain a symbolic operation result Inum_SIGN; (4) performing Fourier inversion on the image processed through a symbolic operator to obtain an IFFT processing result Inum_IFFT; (5) performing Gaussian convolution smoothing, and performing smoothing filtering on the inversion image to obtain a significant image Inum_SIG; (6) performing area selection on the significant image through a threshold to obtain a selected image Inum1; (7) averaging coordinates of the unmanned aerial vehicle area in the Inum1 to serve as the location coordinate (xnum_out, ynum_out) of the unmanned aerial vehicle. The method has the advantages of being simple in principle, easy to implement, high in accuracy and the like.

The invention provides an MSRCR image defogging method based on multi-channel convolution. The method comprises the steps of performing guided filtering processing on a source image, convolving the processed R, G and B channels by six 3*3 Gaussian convolution kernels; obtaining six feature maps with the same size as a single input channel, enhancing the six feature maps corresponding to each channel through a Retinex algorithm, then carrying out linear weighted fusion, carrying out weighted fusion on an image after Retinex enhancement and a detail image after secondary guided filtering processing, and reconstructing a final defogged image. According to the invention, convolution is carried out by using a multi-scale Gaussian convolution kernel; the finer characteristic estimation incidentcomponents are extracted; the multi-scale linear weighted Retinex enhancement is performed on an incident component, and meanwhile, the smooth constraint of the incident component and a reflection image is considered in secondary guide filtering, so that the processed image not only meets the smooth constraint, but also reduces the noise, and the linear weighted fusion is performed on the two enhanced images to realize defogging of the image.

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 relates to an adaptive clustering method of a large number of high-dimensionality image local feature points. The method includes the following steps that: a scale space is constructed for an input image; a Gaussian difference scale space is constructed through using Gaussian difference kernels of different scales and image convolution; each sampling point is compared with all points adjacent to the sampling point; the positions and scales of feature points are determined accurately through fitting a three-dimensional quadratic function; Gaussian filtering is performed on the input image; the variance and Gaussian convolution kernel of a Gaussian filtering function are changed; the feature points are clustered; the value of a bias parameter is selected; the values of the influence degrees and membership degrees of sampling points are calculated; the values of the influence degrees and membership degrees are calculated constantly until an appropriate clustering center is found out; when the number of the times of calculation exceeds a set maximum value, or the clustering center does not change after a plurality of times of calculation, calculation is terminated; and curve fitting is performed on 10 groups of clustering results, the clustering classes of the feature points corresponding to the input image are found out, the feature points of the input image are clustered. With the method of the present invention adopted, the efficiency of image analysis is significantly improved.

The invention discloses a method for monitoring poultry quantity in a breeding farm based on depth learning analysis. The method comprises the following steps: (1) according to a fixed sampling frequency, converting the monitoring video into a static chart to obtain a training picture data set; (2), marking that picture data set; (3), performing Gaussian convolution on each labeled picture to convert it into a density map, and the actual quantity of poultry in each picture is calculated; (4), dividing the picture data set at a ratio of 8:2 into training set and test set as the input of training model and test model respectively, training depth learning model offline(as shown in Figure 3), selecting the best model for poultry quantity monitoring by comparing the MAE of different parameter models on the test set; (5), real-time decoding the poultry quantity monitoring video obtained in step (4), inputting into the trained model, integrating the density map output by the model, and obtaining the poultry quantity within the monitoring range. The invention realizes the real-time monitoring of the poultry quantity more accurately.

The invention discloses a local field estimation method based on magnetic resonance dipole field space distribution, and relates to nuclear magnetic resonance quantitative magnetic susceptibility imaging. The method comprises the following steps: estimating a writhing number based on a phase image; estimating a Gaussian convolution kernel standard deviation Sigma (r) modulated by the amplitude, gradient and writhing number of the phase image; dividing an interested region into a high magnetic susceptibility variation region and a uniform susceptibility region according to space information; regulating the Gaussian convolution kernel standard deviation voxel by voxel according to the space information of the phase image, namely the amplitude, gradient and writhing number of the phase image,and further estimating the weight of a sphere center point and other points. The method can effectively remove background field of a strong magnetic susceptibility region in different imaging directions and accurately estimates local field, and meanwhile keeps the completeness of brain tissue, and the effect of the method is obviously superior to that of the existing method.

The invention discloses a deep learning and recognition method for a dense bird flock. The method comprises a probability density map generation process and a training process of a full convolutionalneural network. The probability density map generation process comprises the following steps: inputting bird flock photos into a bird flock photo set; making a color table; dotting and marking all birds in the bird flock photo; converting the bird flock photo into a continuous density function by using Gaussian convolution; mapping the continuous density function with a color table lookup table toobtain a corresponding probability density map A. The training process of the full convolutional neural network comprises the following steps: carrying out image addition processing on a bird flock photo; obtaining a bird flock image; establishing an FCNN full convolutional neural network; obtaining a corresponding loss function; inputting the bird flock image into a neural network to obtain a probability density map B; and calculating a function value of the corresponding loss function to obtain a corresponding weight value. The method has the remarkable effects that the probability-based deep learning technology is used for image training, so that the number of birds is estimated.

The invention discloses a rapid SIFT extraction method based on information quantity. The method comprises the steps that one-time Gaussian convolution is carried out on the basis of an original drawing to obtain a Gaussian convolution drawing; the eight-point neighborhood extremum method is adopted on the original drawing and the Gaussian convolution drawing to obtain candidate feature points, vertical projection of each candidate feature point is carried out on the original drawing, the circular region with the projection point on the original drawing as the center and four pixels as the radius is searched for the candidate feature point, closest to the projection point, on the original drawing, and the found candidate feature point is used as the feature point of the unchanged size; unit radius information quantity obtained by dividing the circular regions with the radiuses ranging from one pixel to twenty pixels by the radius is calculated, and the radius obtained when the information quantity is the maximum value is used as the local size; a twelve-dimensional SSIFT feature vector is calculated, the vector is normalized, and rapid matching among images is achieved. The rapid SIFT extraction method has the advantages that the calculation time is shortened, the real-time performance is improved, and target matching can be carried out on the noisy complex environment.

The invention provides a feature extraction method for identifying and tracking a moving target. The method comprises the following steps: inputting an image and preprocessing the input image; using aGaussian kernel convolution function to generate a scale space pyramid; detecting feature points by adopting a Hessian matrix, roughly positioning the feature points, accurately positioning the feature points, extracting the main direction and the main curvature of the geometric surface of the image, calculating the gradient and the direction of the feature points according to the main curvature,and counting and labeling the obtained gradient amplitude and direction information by utilizing a method similar to HOG. According to the method, through using the Gaussian convolution function to construct the scale space, the scale invariant characteristic of the extracted feature points can be improved. The feature points of accurate positioning can be accurately acquired, gradient amplitudeand direction information of the acquired feature points can be guaranteed, good robustness on illumination can be realized, and good robustness on illumination, rotation, scale, visual angle, non-rigid deformation and other changes is realized.

The invention discloses an image enhancement method based on sine curve change and application thereof, and the method comprises the following steps: obtaining an RGB color image, presetting a brightness increment parameter, and carrying out the sine normalization processing of the RGB color image, thereby obtaining a global sine normalization image; graying the RGB color image to obtain a grey-scale map, carrying out Sobel edge detection, respectively and independently carrying out convolution on the grey-scale map in the x direction and the y direction by adopting a transverse difference operator and a longitudinal difference operator, and carrying out weighted merging to obtain an edge image; and performing Gaussian convolution kernel filtering smoothing on the edge image, performing convolution on a Gaussian convolution kernel and the edge image, presetting a sharpening intensity adjustment parameter, performing exponentiation according to the sharpening intensity adjustment parameter to obtain an enhancement index, and performing exponentiation on the global sine normalized image and the enhancement index to obtain an enhanced image. According to the method, the image brightness increment is in normal distribution, so that smooth adjustment is realized, the image is sharpened to different degrees by adopting the edge sharpening weight, and a better image enhancement effect is obtained.

The invention discloses a quick defogging processing assembly for fog images. The quick defogging processing assembly for the fog images comprises a video input checking module, a video input statistic module, a video processing module, a video output module and an external storage controlling module, wherein the video processing module is the core of the quick defogging processing assembly and comprises a first-frame removing unit, a frame frequency controlling unit, a pixel minimum extracting unit, a local minimum extracting unit, an image edge expanding unit, a row filtering unit, parallel column filtering units and an enhancement processing unit. By utilizing a quick Gaussian convolution algorithm, the quick defogging processing assembly for the fog images is simple in structure and algorithm, low in realization difficulty and capable of greatly saving sources and reducing delays; by utilizing the multiple parallel column filtering units for row and column filtering of video streaming, the quick defogging processing assembly for the fog images reads data by rows during processing and then skips by columns, and can acquire multiple columns of data simultaneously and input the same into multiple Gaussian row filtering circuits for parallel processing synchronously, so that lost time for skipping by column addresses is compensated, and instantaneity for defogging processing is guaranteed.

The invention discloses a method for detecting the edge of a pantograph slide plate of a city rail train. At first, that gray image signal of the pantograph slide plate is obtained, and the image signal is convolution with a Gaussian convolution kernel with adaptive size; Then the gradient magnitude and direction of the filtered image are calculated by the gradient operator in the 8 neighborhoodsof 3*3. Finally, the gradient amplitude is suppressed by non-maximum method, and the edge image of pantograph skateboard is obtained by double threshold adaptive extraction and edge connection. The invention is based on an improved Canny operator, and has obvious detection result and strong applicability.