33results about How to "Preserve feature information" patented technology

The invention discloses a denoising method for a partial discharge ultrahigh-frequency signal of a transformer based on the improved variational mode and SVD (singular value decomposition). The methodcomprises the following steps: collecting a partial discharge signal of oil-paper insulation of the transformer by using an ultra-high frequency detection method, performing the variational mode decomposition of the collected partial discharge signal; optimizing the parameters of the variational mode decomposition algorithm through using the evolutionary algorithm, and introducing a kurtosis indicator to filter out narrowband noise; performing the SVD of a signal with the narrow-band noise being filtered; improving a singular value screening process in the SVD algorithm through using a clustering algorithm, thereby achieving the better filtering of white noise. The method has the characteristics of high precision, good denoising effect and small distortion, and is suitable for real-time monitoring of oil-paper insulation of the transformer, partial discharge signal detection and the like.

The invention discloses a method for denoising images of shaft walls of deep shafts, comprising five steps: Step 1, constructing a denoising model; Step 2, designing a loss function; Step 3, using standard images to train the denoising model to obtain model parameters; Step 4 , use the actual image to train and test the denoising model, modify the model parameters, and obtain the ELU-CNN denoising model; Step 5, input the noisy well wall image into the ELU-CNN denoising model to obtain the denoising result. The denoising model is 28 layers deep, including 5 feature extraction modules FEM and skip links. The skip links combine the output features of the first convolutional layer with the output features of each FEM in series, so that the low-level features extracted by the first layer are continuously used. , to ensure the full extraction of image features. The invention can well preserve the texture features of the damaged part of the well wall when removing the blind noise of the well wall image.

The invention discloses a bank note one-dimensional signal detection method and device. The method comprises: converting bank note one-dimensional signal data into a two-dimensional gray value; preprocessing the two-dimensional gray value; and detecting the two-dimensional gray value through a two-dimensional data processing method, so that it can be detected whether an adhesive tape is bonded on the bank note, the authenticity and face value of the bank note. The present invention solves the problems in the prior art that, in the existing one-dimensional signal detection, abnormal samples are difficult to be distinguished, and normal samples are falsely denied, and bank notes bonded with adhesive taps are difficult to be identified accurately.

The invention discloses a method of automatically analyzing a 3D point cloud registering error, based on stereo grids. The method of automatically analyzing a 3D point cloud registering error, based on stereo grids includes the following steps: 1) obtaining a source point cloud; 2) obtaining a transformation matrix; 3) performing down-sampling on the source point cloud; 4) filtering the source point cloud; 5) obtaining the target point cloud; 6) extracting the characteristic points; 7) calculating the characteristic vector of the characteristic points; 8) looking for the corresponding points between the source point cloud and the target point cloud; 9) calculating the average distance between the corresponding points; 10) when the number of the discovered corresponding points is reduced to be less than the threshold, entering the step 11, or modifying the control variable looking for the corresponding points, and returning to the step 8; and 11) through weighted average, obtaining the error of the laser radar 3D point cloud registering. The method of automatically analyzing a 3D point cloud registering error, based on stereo grids can quantitatively compare the accuracy of using different registering algorithms to construct a 3D scene model, can help the registering algorithm to determine the suitable point cloud scale in practical application, and can be used to the laser radar three dimensional plotting and structural measurement field with higher requirement for registering accuracy.

The invention discloses a geologic body data visualization oriented feature-preservation quantifying method. The geologic body data visualization oriented feature-preservation quantifying method includes smoothing initial geologic body data by a bilateral filtering process, respectively performing statistics on a histogram of the initial geologic body data and a histogram smoothed geologic body data to obtain a difference histogram of the two histograms, and acquiring high-frequency detail feature intervals by a Gaussian fitting process; defining a local correlation computation process, performing statistics on local correlation histograms of the initial geologic body data and acquiring feature intervals with high local correlation in the geologic body data by the Gaussian fitting process; removing the high-frequency detail feature intervals and local correlated feature intervals in high-dynamic-range intervals of the initial geologic body data to acquire singularity feature intervals in the initial geologic body data. The geologic body data visualization oriented feature-preservation quantifying method has the advantages that a user can interactively define significance proportions of features according to requirements of the user, the initial geologic body data in a 32-bit high-dynamic range can be quantified into 8-bit geologic body data by the aid of a segmented quantification scheme, and accordingly the initial geologic body data can be quantified in a feature-preservation manner.

The invention provides a hyperspectral remote sensing image classification method based on the attention mechanism and convolutional neural network, which performs principal component analysis on the original hyperspectral remote sensing image to reduce the dimensionality, and samples and blocks the hyperspectral data after dimensionality reduction; then performs 3D Convolution operation and pooling operation to obtain the intermediate feature map; then each spectral vector of the intermediate feature is multiplied by the spectral attention module and each spatial feature and the spatial attention module respectively to obtain the attention enhancement sample; Then perform another convolution operation and attention enhancement operation; then input the intermediate feature map obtained through the 3D convolution operation into the classifier for classification. The invention has the beneficial effects of reducing the classification cost, improving the classification performance, realizing adaptive feature refinement through the extraction and enhancement of sample features, and further improving the classification accuracy of hyperspectral remote sensing images.