148 results about "Scale invariance" patented technology

The invention disclosers a human body behavior recognizing method based on an RGB-D video and belongs to the technical field of computer vision behavior recognition. The method includes extracting the dense Moving Pose feature, the SHOPC feature and the HOG3D feature from the RGB-D video acquired from an RGB-D camera according to the principle that human body behaviors of different classes in the RGB-D video have different moving information, geographic information and texture information, adopting an edge-limited multi-core learning method to conduct feature fusion on the three types of features, and finally adopting an Exemplars-SVM linear classifier is adopted to judge human body behavior action. Compared with the prior art, the three types of features extracted have the advantages of illumination invariance, scale invariance and view angle invariance, obvious robustness is achieved for the appearance difference and the behavior action process difference of action executers , and the human body behavior action recognition accuracy is improved to some extent.

The invention discloses a text detection and recognition method combining character level and character string level classification. According to the method, pixel sets possibly belonging to the same character are extracted from images to form alternate characters; alternate characters which do not meet the geometric feature statistic rule are filtered out; a character level classifier based on the character rotation and dimension invariance features is adopted for classifying the alternate characters, and the probability that the alternate characters are certain characters is determined; the characters are combined two by two, and initial character strings are formed; the similarity between every two character strings is calculated, two character strings with the highest similarity are combined into new character strings until no character strings capable of being combined exist; a character level classifier based on the character string structure feathers is adopted for classifying the character strings to determine the character strings with semanteme; and the probability that character strings to be recognized are certain characters is utilized for recognizing the character strings, and the semantic text is obtained. The text detection and recognition method has the advantages that the text detection and recognition process is used as a whole, the interaction of the text detection and the recognition is utilized for improving the result precision, and simplicity and high efficiency are realized.

The invention discloses an improved SURF fast matching method, and belongs to the technical field of digital image processing. The improved SURF fast matching method comprises the steps that an original image and an integral image are converted and detection of feature points is conducted through a Hessian matrix; the feature points, obtained by using a scale space, are scale-invariant; positioning of the primary direction of the feature points are conducted by calculating the maximum value responded to a Harr small wave; the feature points are classified through an improved feature-point classifying method and feature describing is conducted on the feature points and a 66 dimension feature vector is formed; matching of feature points in a set is conducted finally. According to the improved SURF fast matching method, selected feature points in one image are not needed to be matched with all the feature points in another image, both matching speed and accuracy are greatly improved, and the improved SURF fast matching method has the advantages that the improved SURF fast matching method has properties, such as scale invariance and translation rotating resistance, of classic SURF (Speeded Up Robust Features).

The invention provides an icon recognition method and device. The method comprises the following steps: a to-be-recognized image containing a specific icon is acquired; an image processing method is used to process the to-be-recognized image to detect the feature points of the specific icon, and feature vectors of the feature points are extracted to form a unique feature vector group; the generated feature vector group and a feature vector group uniquely corresponding to reference icons preset in a feature database are matched one by one to determine a target icon matched with the specific icon; and the icon information corresponding to the target icon is acquired from a third-party database and is outputted. The method carries out icon recognition based on an SIFT algorithm with scale invariance, the icon recognition precision is enhanced, and robustness of the scheme towards rotation, brightness change and scale transformation happening to the image are enhanced.

The invention relates to a lightweight remote sensing target detection method based on SEYOLOv3, which belongs to the technical field of target detection, and comprises the following steps: 1, takinga YOLOv3 algorithm as a basic model framework, and in order to reduce network parameters and improve network reasoning speed, designing a lightweight trunk feature extraction network; 2, in order to improve the scale invariance of the features and reduce the over-fitting risk, a spatial pyramid pooling (SPP) algorithm is provided, and pooling of three scales is carried out to obtain an output feature vector with a fixed length; a spatial attention model SE module is introduced, useless information is further compressed, and useful information is enhanced; and 3, updating parameters through iterative training to obtain a final network model, adopting multi-scale prediction by utilizing the model, and predicting a final result through detection heads of three scales. According to the method,while the reasoning speed of the network is effectively improved, the precision is ensured, the feature expression capability of the network is enhanced, and the scale invariance is improved.

The invention discloses a free-hand sketch offline identification and reshaping method. The method comprises the following steps of: preprocessing an input image; converting a discrete and disordered point set in each connected domain into an ordered point sequence for compression; fitting a plurality of straight lines of the point sequence by using a dynamic programming algorithm, and determining the number of optimum fitted straight lines to obtain the strokes, represented by the straight lines, of each connected domain; analyzing a stroke result obtained by fitting the plurality of straight lines; if the number of the fitted straight lines is greater than the maximum edge number of a shape identified by a system, performing order reduction on the strokes, classifying the strokes, and calculating distance between the strokes; and selecting the nearer strokes for combination and analysis, and performing verification according to the geometrical characteristic to determine the shape formed by combination of the input strokes. The method is higher in identification rate, and the shape to be identified has scale invariance and rotation invariance; and the algorithm supports a multi-stroke form for identifying limited shapes, and the problems in the identification completely based on the geometrical characteristic are solved.

The invention discloses a shape matching and target recognition method based on a PCA-SC algorithm. The method comprises the steps of carrying out preprocessing on a target image, filtering part of noises in the target image, extracting the edge of the target image, extracting information of boundary contour points, working out the rectangular coordinate parameters of the contour points, converting the contour points from rectangular coordinates into polar coordinates, obtaining a corresponding logarithmic polar histogram of each point to forming a local feature descriptor, forming a covariance matrix, extracting a corresponding feature vector of a larger characteristic value of the matrix, adopting a linear transformation method to drop the matrix from high dimension to low dimension, forming a new characteristic matrix, wherein the new characteristic matrix is used for the shape matching and the target recognition, calculating matching degree, and obtaining a matching degree value between the target image and each template image. According to the shape matching and target recognition method based on the PCA-SC algorithm, characteristic extracting and effective representation for the image can be achieved, scale invariance, rotation invariance and translation invariance are achieved, accuracy rate and efficiency are improved, and interference of the noise is effectively restrained.

The present invention discloses a three-dimensional scene reconstruction method and device based on vision SLAM. The method comprises the following steps of: vision information obtaining: allowing a mobile robot to freely move in a three-dimensional scene to collect images, and employing an SLAM algorithm to estimate pose information of the mobile robot; point cloud reconstruction: employing an SIFT algorithm to extract feature points in the collected images for matching, employing the SFM algorithm to perform sparse point cloud reconstruction of the matched feature points and the corresponding mobile robot pose information, and performing dense point cloud reconstruction of the reconstructed sparse point cloud; and surface reconstruction: performing surface reconstruction of the reconstructed dense point cloud to complete reconstruction of the three-dimensional scene. The three-dimensional scene reconstruction method and device are high in operation speed, low in requirement for hardware and good in scale invariance. Besides, the three-dimensional scene reconstruction method and device further reconstruct object surface information so as to improve the precision of the reconstructed three-dimensional scene.

The invention discloses a feature extraction and dimension-reduced neural network-based visual SLAM closed-loop detection method. According to the feature extraction and dimension-reduced neural network-based visual SLAM closed-loop detection method, a convolutional neural network model is trained through a large number of data set to endow a network with feature learning capacity, so that similarity comparison between images can be converted into similarity comparison between feature vectors; for further improving the detecting speed, the last layer of the convolutional neural network is provided with an auto-encoder network to dimension-reducing extracted image features; the convolutional neural network has the advantages of translation invariance, scale invariance and the like and accordingly can effectively overcome the shortcoming that traditional artificial feature extraction is sensitive to environmental change and achieve a higher feature extraction speed. The feature extraction and dimension-reduced neural network-based visual SLAM closed-loop detection method can overcome the shortcomings of short feature extraction time and large influence of environmental change and light change in traditional visual SLAM closed-loop detection methods, effectively improve the accuracy and recall rate of closed-loop detection and achieve high significance to structuring globally uniform environmental maps.

The present invention provides a super-resolution reconstruction method for an image. The super-resolution reconstruction method comprises: acquiring a plurality of low-resolution images and performing image registration; performing non-sampling morphological wavelet decomposition on each image in a registered image sequence to obtain corresponding low-frequency coefficient sequence and high-frequency coefficient sequence; performing interpolation on the low-frequency coefficient sequence and the high-frequency coefficient sequence separately, and fusing the amplified low-frequency coefficient sequence and the amplified high-frequency coefficient sequence to obtain a fused coefficient; and performing non-sampling morphological wavelet inverse transformation on the fused coefficient to obtain a reconstructed image. With the adoption of the reconstruction method, the scale invariance and the multidirectional property can be kept, a reconstruction calculation process is simple, and the quality of the reconstructed image is better.

The invention provides a method and a system for quickly matching images on the basis of feature states and global consistency. The method includes utilizing detection angular points as to-be-matched feature points and identifying feature neighborhood states of the feature points by the aid of state templates; utilizing the feature points as centers and computing main directions of descriptors; turning the main directions, describing feature neighborhoods of the feature points, combining binary texture features and statistic features with one another and creating RBT-OGMH feature descriptors; matching the images for two types of different descriptors by the aid of different types of similarity measurement; quickly determining transformation matrixes by the aid of consistent features of spatial domains on the basis of error fluctuation amplitude minimization clustering, and eliminating error matching point pairs to obtain ultimate correct matching point pairs. According to the technical scheme, the method and the system have the advantages that the method and the system have rotational invariance and certain scale invariance, and problems of image blurring, illumination change, low contrast and image deformation can be effectively solved; the method and the system are high in matching speed and precision as compared with the prior art.

The invention discloses a pig water drinking behavior identification method based on contours. The method comprises steps that firstly, water drinking areas are separated from acquired look-down grouped pig frames, a two-dimensional OTSU method is utilized to acquire the preliminary segmentation result, and binary and morphological processing on the preliminary segmentation result is carried out to acquire the target contours; secondly, a polygon approximation method is utilized to acquire polygon fitting maps of the contours, for the polygons, two-dimensional characteristic quantities having scale invariance and rotation invariance are extracted; thirdly, optimal matching of the two polygon relevant characteristic quantities is carried out through a Hungary algorithm; lastly, similarity between the two polygons is calculated, matching of to-be-identified contours and training sample contours is accomplished, and pig water drinking behavior identification is realized. The method is advantaged in that bases are established for identification researches on feeding and defecating behaviors of the group pigs, and the new way is provided for exploring behavior identification of the animal husbandry.

The present invention relates to an ORB feature point matching method with scale invariance. The method is characterized by comprising the steps of a step S1 of inputting an image to be detected, performing improved SURF feature point detection on the image, and determining coordinates of the feature points; a step S2 of establishing an image pyramid for the image in the step S1; a step S3 of removing the feature points close to edges of the image; a step S4 of calculating directions of centers of mass of the remained feature points; a step S5 of calculating ORB feature point descriptors; a step S6 of adopting a K-nearest neighbor algorithm to carry out feature point matching; and a step S7 of screening feature point matching pairs and outputting the detected image. According to the method provided by the present invention, the SURF with the scale invariance and the ORB are combined, the image pyramid is introduced, and an ORB feature point matching algorithm is improved, so as to enable the method to have the scale invariance and maintain the characteristic of fastness of the ORB algorithm.

The invention discloses a gradient binaryzation based rotation-invariant and scale-invariant scene matching method, and relates to the field of scene recognition. According to the method, on the basis of a classical binary description BRIEF algorithm in which only gray scale intensity is compared, horizontal and vertical gradient comparison is added, texture information of a sampled area is saved, and accordingly, matching error rate is reduced. Moreover, an image scale pyramid is created, image feature point detection and feature description are performed within different scales, gravity center vector directions are added during descriptor calculation, and direction and scale invariance of binary descriptors is achieved. Experiments show that binaryzation based rotation-invariant gradient sampling descriptors have high robustness, and the matching accuracy rate is 73.06% higher than that of the BRIEF algorithm in average when a scene image is rotated greatly and the scale is varied.

The present invention relates to a contour chord angle feature based identification method for a blocked target. The method comprises: establishing a template library of local features of a plurality of target images; extracting a contour feature of a target edge; constructing a chord angle feature descriptor of each contour point; describing a blocked contour by using a self-containing attribute of the chord angle feature descriptor, to obtain a chord angle feature description matrix of contour segments; calculating distances between the chord angle feature descriptors of target image contour points and the chord angle feature descriptors of the contour points of the local feature in the template library by using an L1 measurement method, to obtain a matching cost matrix; and calculating a similarity of the matching cost matrix by using an integral graph algorithm, so as to identify a partially blocked object. According to the present invention, a contour space position feature of a target shape can be extracted, and blocked targets can be identified, and scale invariance, rotational invariance and translation invariance are achieved, thereby increasing the accuracy and robustness of target identification and shape search.

The invention discloses a monocular visual mileometer positioning method. The method comprises the following steps that an input image of a present frame is read; a parallax based method is used to initialize a visual mileometer, a subsequent motion estimation step is started after success of initialization; a nonlinear optimization problem by taking pose as an optimization variable is formed fora last frame on the basis of gray scale invariance, and an initial inter-frame pose is obtained; a local map is constructed on the basis of the initial pose, and a density based tracking strategy is used to complete projective point characteristic matching and sub-pixel position optimization of the local map; and a more accurate constraint relation is obtained by tracking of the local map, the relation further optimizes the pose and map points, and a final positioning result is output. The monocular visual mileometer positioning method can reach higher positioning precision, and can reach therobustness and instantaneity needed by unmanned aerial vehicle navigation applications.

The invention relates to a solar image registration method based on normalization cross-correlation and SIFT, and belongs to the field of astronomical technology and image processing. The method comprises the following steps: firstly, carrying out downsampling preprocessing on a group of full-solar observation images from a space-based (SDO) and a group of solar local high-resolution observation images from a ground telescope (NVST) respectively; secondly, using a normalization cross-correlation matching algorithm to solve the problem that the view fields of the to-be-registered area are inconsistent; calculating the optimal matching position of the full-day image and the local high-resolution image, and using the position to intercept a sub-image to serve as a to-be-registered reference image; Performing feature detection on the to-be-registered image by adopting a feature detection operator (SIFT) based on scale invariance to obtain a feature point set; Using an MLESAC algorithm to eliminate mismatching feature point pairs; And finally, solving transformation parameters among the images by using a least square method to obtain a registration result. According to the invention, rapid, automatic and high-precision registration of solar images of different observation sources is realized.

The invention relates to a method for tracking a point feature based on fractional-order differentiation, comprising detecting a point feature by adopting a method based on the fractional-order differentiation; forecasting the location of the next frame point by using a Kalman method or an extension method; searching in a given area, carrying out a similarity measurement, and acquiring a corresponding tracking point if requirements are satisfied; otherwise, considering the corresponding tracking point to be absent, and for such a point, considering the tracking is lost if a corresponding matched tracking point is still absent in a range of following k frames, wherein the value of k is larger than 2; and updating the point feature if the tracking is normal. The fractional-order differentiation has advantages over integral-order differentiation in presenting areas which have abundant texture details and inconspicuous texture information. Different differential gradient images are formed for fractional-order differentiation with different directions and different orders, and convolution directional diagrams with different scales are formed by combining the different differential gradient images with Gaussian kernel convolution in different sizes respectively, so that significant changes presented by the point feature are ensured when the direction is changed, having properties of rotation invariance, translation and scale invariance.

The invention provides a typical ship target identification method based on graded invariance features. At first, the binary entropy and the normalized inertia moment of each image ship target are extracted as the primary feature; and then each image is subjected to wavelet decomposition to form four sub-images, and the weighted Hu moment, Zernike moment, and Fourier descriptor of the each sub-image ship target are extracted to be the secondary feature; the polar coordinate shape matrix of each image ship target is taken as the trinary feature; and all of the features are modified to have the properties of translation, rotation, and scaling invariance. The experimental result of a recognition classifier shows that the algorithm can describe the typical ship targets in satellite remote sensing images in details step by step, and the recognition accuracy is high. The method can be applied to the typical ship target recognition of the satellite remote sensing image database, and is an engineering method that is high in universality.

The invention discloses a method for constructing a color background of a grayscale target image, comprising: adopting an infrared camera to capture a plurality of color scene images; obtaining a panorama of the fixed region through image stitching technology; then acquiring a part of the grayscale scene graph in the panoramic information by using the infrared camera; performing pixel-scale scaling on the same object in the panorama and grayscale scene graph, so that the size of the grayscale scene graph is scaled to the same size as the corresponding position of the panorama; and finally, obtaining the color background image at the same position as the grayscale scene image from the panorama by template matching. The invention avoids the information distortion problem caused by directly coloring the color image, and may highly restore the scene information. The invention utilizes the scale invariance of the SIFT features in the prior knowledge to obtain a good detection effect. The invention obtains a color background image skillfully by using the image processing method, which is less in time consuming, low in cost and high in precision.

The invention provides a monocular vision mileage measuring method and an odometer based on image characteristics. The method comprises the following steps: (1) calibrating a camera; (2) calculating 2D characteristic points of two adjacent frames of images along the advancing direction; (3) matching the 2D characteristic points to find corresponding characteristic points in the two frames of images; (4) calculating the 3D coordinates of the corresponding characteristic points in the two frames of images, and calculating the pose of the camera according to the 3D coordinates and the 2D coordinates of the corresponding characteristic points to obtain the relative displacement of the camera; and (5) performing the same operation on the subsequent frame, and finally accumulating all the displacements to obtain the mileage. Compared with a binocular vision-based method, the method for measuring the mileage by monocular vision is simple in equipment and low in cost; and compared with a method based on sift and Harris angular points, the method has the advantages that the image characteristic calculation speed is higher, and the rotation scale invariance and the real-time processing can be realized.

The invention discloses a non-local means filtering method for a speckle noise pollution image. The method comprises that iterative computation of a neurons firing state image series of the speckle noise pollution image is conducted through a pulse transmission cortex model, a Renyi entropy vector is extracted by the neurons firing state image series, and non-local means filtering is conducted on the speckle noise pollution image based on the Renyi entropy vector, so that a denoised gray value is obtained. By the aid of the method, rotation invariance, translation invariance and scaling invariance can be extracted from the image containing speckle noise, the method can use more image information for denoising than traditional methods, besides, the similarity between two image pixel blocks can be calculated reasonably, image noise can be suppressed significantly, and the peak signal-to-noise ratio of the image can be improved, so that detailed information of the image can be effectively protected.