50 results about "Gaussian image" patented technology

A depth image partitioning and agent geometry based virtual and real collision interaction method and apparatus are disclosed. The application discloses a method for realizing collision interaction of a virtual object and a real object by partitioning a scene in a depth image and constructing agent geometries based on a partitioning result. Firstly, color image information and depth image information of a current real scene are obtained; secondly, normal information in the depth image and connectivity are taken as criterions, the depth image is subjected to regional partitioning with a regional growth method, and on this basis, a main plane region in the scene is identified by utilizing a plane judgment technology of a Gaussian image; thirdly, other point cloud regions except the main plane region are subjected to fusion processing to obtain other main object regions in the scene; fourthly, constructing a virtual plane for an identified main plane and constructing a bounding box for a partitioned object; and finally, superposing the agent geometries onto the real object, endowing the real object with physical attributes, and realizing collision interaction of the virtual object and the real object by simulation. The depth image partitioning method disclosed by the application is quick, effective and more realistic in virtual and real interaction effect.

A robust method for extracting a two-dimensional code area in an image is disclosed, which comprises the following steps of: firstly, building a multi-scale Gaussian image pyramid according to an original image; secondly, partitioning each layer in the image pyramid; thirdly, binarizing all the image blocks, respectively, in order to obtain the binarization result of each layer of image; fourthly, fusing the binary images under a plurality of scales to obtain an image partition result; and finally, searching for the two-dimensional code area in unit of the image block and extracting the two-dimensional code area in the image by analyzing and calculating the convex hull of a characteristic point set through a connecting body. The method disclosed by the invention is capable of handling the common complex cases such as uneven illumination, background interference and the like, and has good robustness; as a search policy from coarse to fine under a plurality of scales is utilized, the method is simple and quick, and thereby is capable of taking both of the instantaneity and accuracy of handling into account.

The invention relates to a multi-scale spatial projecting and remote sensing image fusing method. The method comprises the following steps of: calculating the brightness image of a primary low-spatial resolution multispectral image by using the primary low-spatial resolution multispectral image, respectively generating the Gaussian image cube of the brightness image and a primary high-spatial resolution panchromatic image through improved Gaussian scale-space theory, extracting spatial detail characteristics of the brightness image and the panchromatic image through the difference between layers of the Gaussian image cube, and projecting the spatial detail characteristics onto the primary multispectral image according to a weighted fusion scheme to obtain a fused multispectral image with high spatial resolution and high spectral resolution. By the method, the spectral information of the primary multispectral image is well kept when the spatial detail performance ability of the fused image is improved; the physical significance is clear; the realization structure is simple; and the fusion effect is good.

The invention discloses a gait recognition method based on a depth neural network, and the method comprises the steps: carrying out the preprocessing of an original gait video image, extracting the gait Gaussian image of the original gait video image, and dividing a sample data set into a training set and a testing set at a proportion of 5: 1 according to a designed rule; building an eight-layer convolution neural network model consistent with an AlexNet structure, modifying the number of neurons of the last layer of the model, and enabling the model to adapt to a gait classification task; Carrying out the initialization of the parameters of a trained AlexNet model through the former seven layers of the model, and carrying out the random initialization of the last layer of the model; training the convolution neural network through the training set, and enabling the convolution neural network to effectively complete the gait recognition. The method is higher in robustness, can achieve the more effective recognition of the identity of a person under the condition of various types of concomitant variables, and can remarkably reduce the calculation resources and time of a training model.

A method for enhancing an image is provided. The method generates a plurality of Gaussian images from the image, and generates a first plurality of Laplacian images from the generated Gaussian images, the plurality of Laplacian images constituting a Laplacian pyramid of images. The method generates a histogram for each of the plurality of Laplacian images, and shapes each of the generated histograms using a predetermined probability density function to produce a second plurality of Laplacian images. The method further combines iteratively each of the second Laplacian images with each of the plurality of Gaussian images expanded at each iteration to generate an enhanced image.

The embodiment of the invention discloses a method and a system for tracking vehicles. The method comprises the following steps of: establishing a gaussian image pyramid according to a video image of the current frame, and performing FAST feature detection on the gaussian image pyramid to obtain a first FAST feature set; acquiring a dot pair sequence set corresponding to the first FAST feature set according to the first FAST; generating BRIEF descriptors of feature points in the first FAST feature set according to the dot pair sequence corresponding to the feature points to obtain a descriptor set of the feature points of the current frame; establishing a feature dot pair sequence according to preset tracking radius, the descriptor set of the feature points of the current frame and the descriptor set of the feature points of a tracked target in the previous frame; and determining the position of the tracked target in the current frame according to the feature dot pair sequence. By the method and the system, the tracking of the vehicles can be realized effectively, the lost of the tracked target is avoided, and a tracking effect is good.

The invention provides a parallel and adaptive matching method for acquiring remote sensing images with homogeneously-distributed matched points. The method comprises the following steps of: deblocking overlapped regions of the remote sensing images with different resolutions; parallelly extracting features of matched images aiming at each sub block; during extraction, automatically controlling the logarithmic number of matched points through the classification of a Gaussian image tower with different layers and a Gaussian differential image, so that the homogeneously-distributed matched points of the remote sensing images are parallelly and adaptively acquired, and the matching of the remote sensing images with different resolutions is achieved. The method can realize the parallel and efficient matching of the remote sensing images and can also realize the homogeneous distribution of the matched points.

The invention discloses a device and method for measuring the curvature radius of a spherical mirror based on a pinhole type point diffraction interferometer. Double quantitative defocusing is realized through sequentially inserting two parallel flat plates with different thicknesses in an interfering cavity of the pinhole type point diffraction interferometer. When the parallel flat plates are not placed in a test optical path, wavefront data W0 are obtained through phase shifting measurement. After the two parallel flat plates with different thicknesses are respectively placed in the test optical path, two different wavefronts W1 and W2 are obtained through phase shifting measurement, and the calculating formula of the curvature radius of the spherical mirror is derived through the difference between the defocusing coefficients of the W1 and the W0, the difference between the defocusing coefficients of the W2 and the W0 and a Gaussian imaging formula. When the non-contact type measurement method is adopted, the damage of the surface of the spherical mirror is avoided, a feasible method is provided for the non-destructive measurement of the curvature radius of the spherical mirror with a large numerical aperture. Meanwhile, the method is suitable for the surface shape test of the spherical mirror with the large numerical aperture.

The invention discloses an image processing method for chest X-ray DR (digital radiography) image rib inhibition. The method comprises the following steps: acquiring a chest X-ray DR image; performing pyramid decomposition on the DR image, performing a down sampling process to obtain a Gaussian image pyramid S, and performing an up sampling process to obtain a Laplacian image pyramid difference chart D(S); taking the minimum S as a current to-be-processed image I; performing filtering processing on the image I by using an adjustable Gabor filter bank so as to obtain a reconstructed image R; differencing the to-be-processed image I and the reconstructed image R to obtain a processing result image E with weakened segment-shaped textures under the scale; and doubling the processing result image E, adding the processing result image E and a corresponding Laplacian image pyramid difference chart D(S) under the size together, and repeating the processing procedure until the size is the same as that of an original DR image, thereby obtaining an image after rib inhibition. According to the method disclosed by the invention, the visual saliency of pulmonary shadows is improved, the workload of doctors is reduced, automatic processing can be realized, and an analysis conclusion is more objective and stable.

The invention provides a robot vision positioning method based on feature point detection and mismatching screening. The robot vision positioning method comprises the following steps: A, constructinga Gaussian image pyramid; B, performing FAST feature point detection on each layer of image by using a quadrilateral model; C, constructing a BRIEF descriptor for the extracted feature points based onneighborhood pixel values; D, conducting violence matching based on the BRIEF descriptor; and E, screening mismatched feature points, and corresponding the matched feature points to an original reference image and a to-be-matched image based on a screening result. The method has the advantages that the feature points are detected in a quadrilateral mode, the detection speed is improved, the defect that the FAST feature points do not have scales is eliminated by constructing a Gaussian image pyramid, and based on violent matching, the feature points can be matched with each other, then the mismatched feature points are screened, and a correct matching result is reserved, so that pose calculation is facilitated.

The invention relates to the field of computer vision, in order to optimize the feature descriptor extraction part of the SIFT algorithm, improve the hardware execution speed of the part, maintain thestability of the feature descriptor, and reduce the consumption of FPGA hardware resources. To this end, The technical scheme adopted by the invention is as follows: A low-complexity and fast SIFT feature extraction method based on FPGA utilizes an upsampling module to collect image data to an initial two-dimensional Gaussian filter module G0 to generate a reference Gaussian image, and then utilizes four parallel two-dimensional Gaussian filters G1 to G4 to generate four-layer Gaussian images to form a Gaussian pyramid. Gaussian difference images D1 to D3 obtained by subtracting two adjacentGaussian images constitute a Gaussian difference pyramid. The invention is mainly applied to image feature extraction occasions.

The invention relates to an off-line analysis method for machining allowance of a free-form surface part, which is proposed to solve the disadvantage of low precision of the existing registration method. The method comprises the following steps: obtaining a translation vector capable of making the center of mass of a blank and the center of mass of a design model translate to coincidence; obtaining a translation vector capable of making the center of mass of the blank translate to coincidence; all surface normals of the blank and the design model being mapped to the extended Gaussian image spheres, respectively, and 3D attitude angle combinations capable of making the extended Gaussian image of the blank and the extended Gaussian image of the design model reach the preset correlation conditions being obtained by mapping the blank and the surface normals of the design model to the spheres. The rotation matrix and translation of the blank and the design model can satisfy the full envelope condition, the datum alignment condition and the margin uniformity condition. The invention is applicable to the registration method in the process of machining curved surface parts.

The invention relates to human body detection technology, and discloses a human body detection method based on depth information. The human body detection method settles the problems of severe external environment interference and low applicability in human body detection. The human body detection method comprises an image preprocessing step and a human body detection step, wherein the image preprocessing step comprises the steps of extracting image depth information; performing Gaussian image pyramid down-sampling on a depth image; performing gradient histogram processing on the down-sampled image; performing integrogram calculation on a characteristic channel after gradient histogram processing. The human body detection step comprises the steps of finding out a detection area through interframe motion information; performing characteristic judging on a to-be-detected area by means of a classifier based on a discrete AdaBoost algorithm, and detecting whether the human body exists in the area; and merging windows which are judged as the human body through the classifier, and marking the human body area on the image. The human body detection method is suitable for VR.

The invention discloses a diagnosis method and system based on machine faults. The diagnosis method includes the steps of acquiring each to-be-diagnosed observation signal of the machine faults; according to the observation signals, determining a time-frequency gray image corresponding to each product function; according to the time-frequency gray image corresponding to each product function, determining a time-frequency image pixel matrix; performing redundancy cancellation processing on the time-frequency image pixel matrix to obtain a redundant cancel image pixel matrix; according to scaleinvariant features, determining a Gaussian image difference function, a gradient magnitude and a gradient direction according to the redundant cancel image pixel matrix; determining several base images according to the Gaussian image difference function, the gradient magnitude and the gradient direction; determining image slices according to the base images; inputting the image slices into a support vector machine classification model to achieve fault classification diagnoses. The accuracy of the fault diagnoses is improved.

The invention relates to visual inspection, in particular to a manufacturing method of a visual inspection defect library for practical training. The manufacturing method comprises the following steps: S1, photographing to obtain a defect image; S2, performing fft speed optimization; S3, processing the defect image through a Gaussian filter to obtain a Gaussian image; S4, taking a gray value; S5,performing iterative operation on the difference image; S6, performing fast Fourier transform on the image after the iterative operation; S7, performing inverse transformation on the picture after Fourier transformation; S8, performing threshold segmentation on the transformed image, and taking an extreme value to obtain segmented regions; S9, communicating regions, and selecting regions accordingto specified morphological characteristics; and S10, drawing a circle at the defect part of the area, and calculating the central position of the area. According to the manufacturing method of the visual inspection defect library for practical training provided by the invention, the defect library can be conveniently established for the defects of the workpiece, the universality is high, and practical training personnel can know the specific process.

The invention relates to the technical field of image processing, in particular to a hull surface corrosion degree evaluation method based on image processing, which comprises the following steps: acquiring a surface image of the bottom of a hull, and acquiring a Gaussian pyramid and a Laplacian pyramid according to the surface image, and obtaining enhanced images of the first image enhancement, the second image enhancement and the third image enhancement corresponding to each layer of Gaussian image, and fusing the three enhanced images corresponding to the Gaussian image to obtain a final enhanced image. According to the final enhanced image corresponding to the last layer of Gaussian image of the Gaussian pyramid, the corrosion degree of the hull surface is evaluated, the enhancement effect of the surface image is improved, and the corrosion degree of the hull is accurately evaluated.