42results about How to "Good initial value" patented technology

The invention discloses a three-dimensional point cloud registration method with a low overlap rate, and belongs to the technical field of machine vision. Aiming at the problems of difficulty and low precision in the registration of two point clouds with low overlap rate. First, use the curvature feature of the point cloud to establish a multi-scale descriptor to ensure the integrity of the point cloud data and minimize redundant data; secondly, use the angle difference of the multi-scale descriptor to cluster and block the corresponding relationship to obtain the source point cloud The overlapping area with the target point cloud; finally, the point cloud in the overlapping area and their corresponding relationship are substituted into the convex optimization problem, and the outliers are removed and the corresponding relationship is optimized, and the coarse registration is realized and the ICP algorithm is used for refinement. The invention can narrow the useful search range of point cloud registration, reduce the amount of registration calculation, and provide more advantageous registration accuracy and time efficiency for point cloud data with a low initial overlapping degree. The method can be widely used in three-dimensional model reconstruction, cultural heritage management, robot navigation and positioning and other fields.

Embodiments of the present invention provide an image processing method and device, and an AR apparatus. The method comprises: determining consecutive first and second images in a continuous video frame image sequence, wherein the second image is a current image, and the first image is a previous image of the second image; determining, by using a target region of the first image as a template, location information of the target region in the second image; and performing, by using the location information as an iterative initial value, iterative processing of an entire image alignment on the first image and the second image, wherein the iterative processing of the entire image alignment of the first image and the second image is iterated by using a robust error function. The image processing method and device and the AR apparatus disclosed in the embodiments of the present invention improve success rates in tracking of an entire system.

The invention discloses a seismic super-resolution inversion method based on model-driven deep learning, comprising the following steps: 1) mapping each iteration in the model-driven alternating direction multiplier method ADMM to each layer in the deep network, And use the mathematics-driven method to learn the proximal mapping, complete the construction of the deep network ADMM-SRINet; 2) obtain the label data used to train the deep network ADMM-SRINet; 3) use the obtained label data to perform Training; 4) Inverting the data to be tested using the deep network ADMM-SRINet trained in step 3). This method combines the advantages of the model-driven optimization method and the data-driven deep learning method to make the network structure interpretable; at the same time, due to the increase of physical knowledge, the iterative deep learning method alleviates the requirements for the training set, making The inversion results are more credible.

An image repairing method and a computer-readable storage medium, wherein an image to be repaired is reduced and filtered according to adaptive reduction parameters and filtering parameters, a filtered image to be repaired is obtained, and pixel points to be filled in the filtered image to be repaired are obtained , and obtain a better initial value of the pixel to be filled, based on the constraints that minimize the sum of the divergences of the gradient values ​​of all the pixels to be filled and the initial value of the pixel to be filled, the value of the pixel to be filled is calculated. In the first iterative process, the coarse processing value of the pixel to be filled is obtained, and the coarse processing value of the pixel to be filled is re-used as the initial value of the pixel to be filled. According to the energy equation and the initial value of the pixel to be filled, the pixel to be filled is The second iterative process is performed on the value of the pixel to be filled to obtain the fine-processing value of the pixel to be filled, the fine-processing value of the pixel to be filled is taken as the target value of the pixel to be filled, and the restored image is obtained by up-sampling, which improves the efficiency of image processing. .

The invention discloses a method for improving the expansibility of a deep neural network power flow model. The main steps are: 1) acquiring basic data of a power system; 2) determining a feature vector; 3) establishing an original DNN power flow model; 4) performing an original DNN power flow model Perform training to obtain the original DNN power flow model after training; 5) expand the original DNN power flow model to obtain the expanded DNN power flow model; 6) calculate the probability flow of the expanded system to obtain the probability flow result. The present invention can be widely applied to the probability flow solution of the electric power system, and is especially suitable for the situation that DNN of the original system cannot be applied due to the expansion of the system.

The invention relates to a setting method for the semi-automatic production of a process-capable basic setting for an injection molding machine (1) for processing molding compounds made of plastic. Using an interactive start-up assistant (15) that is integrated in a control apparatus (14) of the injection molding machine (1) by way of software, wherein said start-up assistant (15) guides an operator of the injection molding machine (1) in steps at least through partial sections of the setting process, values or variables for system-relevant process parameters are determined by which a process-capable basic setting is attained as quickly and reliably as possible. For this purpose, prior to starting a test phase, those conditions are defined on a preliminary basis which significantly influence the flow properties of the molding compound, in particular the injection speed and the compound temperature. In addition, a characteristic for the respective flow behavior is determined at least approximately by the start-up assistant (15), in particular a viscosity of the molding compound that is used under said preliminarily established process conditions. Said characteristic for the flow behavior of the molding compound, and optionally other process characteristics which have been determined or specified, are used to establish a provisional, preliminary setting for the target value and / or for the upper threshold value of the injection pressure for the test phase to be started thereafter.

The invention discloses a multi-view point cloud global optimization registration method based on hierarchical closed-loop constraints, which belongs to the intersection field of computer vision and laser scanning data processing. The present invention aims at the spatial difference of the registration primitive neighborhood information generated by different scanning angles of the hand, and analyzes the connectivity of the point cloud according to the feature consistency and geometric consistency of the registration primitive, constructs a connectivity graph for the multi-viewpoint cloud, and converts the The similarity transformation between point clouds is unified to the reference coordinate system. Based on the hierarchical closed-loop constraints, through the error analysis and correction of the least squares joint adjustment, the global optimal fine registration result is obtained.

The invention discloses a breast nodule detection method and system based on ultrasound medicine, which comprises the following steps: acquiring breast ultrasound video data and performing preprocessing, and training the preprocessed video data through a transfer learning method to establish breast nodule detection A model, inputting the video data to be tested into the mammary gland nodule detection model to obtain a result. The present invention provides a breast nodule detection method and system based on ultrasound medicine, using ultrasound breast video image inspection as the main means, aiming to establish an accurate method for judging benign and malignant breast tumors and pathological typing, and solving the problem of dynamic identification and differentiation of breast tumors Difficult clinical problems. The establishment of a highly accurate breast tumor characterization system has great clinical significance for the diagnosis and treatment of breast tumor patients.