32results about How to "Noise robustness" patented technology

The invention discloses a hyperspectral image space spectrum classification method and device considering spectral importance, and the method comprises the steps: calculating the spectral feature importance of a given hyperspectral remote sensing image with hundreds of spectral bands through a random forest; customizing a spectral weight characteristic kernel function, modeling the relative effectof each spectral band in classification by utilizing the extracted spectral characteristic weight, giving a larger weight to the band which is more beneficial to ground object category identification, and improving the spectral discrimination capability of the ground object; constructing spectral unitary potential energy and spatial binary potential energy under a unified framework of a conditional random field, considering spectral characteristic weights in the spectral unitary potential energy, integrating a spectral weight kernel function to improve the distinguishing capability of different spectral characteristics in a nonlinear space, and modeling the spatial correlation of ground objects through the spatial binary potential energy. The beneficial effects of the method are that themethod reduces the impact on classification from an unimportant waveband, improves the discrimination of types, and improves the classification effect.

The invention provides a bird voice recognition technology based on novel noise-proof feature extraction by aiming at the problem of bird voice recognition in various kinds of background noise in ecological environment. The bird voice recognition technology comprises the following steps of firstly, obtaining noise power spectrums by a noise estimation algorithm suitable for highly nonstationary environment; secondly, performing the noise reduction on the voice power spectrums by a multi-band spectral subtraction method; thirdly, extracting anti-noise power normalization cepstrum coefficients (APNCC) by combining the voice power spectrums for noise reduction; and finally, performing contrast experiments under the conditions of different environments and signal to noise ratios (SNR) on the voice of 34 species of birds by means of extracted APNCC, power normalization cepstrum coefficient (PNCC) and Mel frequency cepstrum coefficients (MFCC) by a support vector machine (SVM). The experiments show that the extracted APNCC have a better average recognition effect and higher noise robustness and are more suitable for bird voice recognition in the environment with less than 30 dB of SNR.

The invention discloses a synthetic aperture radar (SAR) target image geometric size estimation method based on orthogonal matching tracking algorithm OMP. (2) obtaining the atomic matrix and the scattering center amplitude vector of the target; (3) obtaining the reconstructed target image; (4) projecting the reconstructed target image onto the coordinate system to generate the target region; (5)obtaining a coordinate vector; (6) rotating the coordinate vector; (7) obtaining the smallest circumscribed rectangle; (8) The size value of the smallest circumscribed rectangle is used as the estimate value of the target geometric size. The invention is universally applicable to stationary radar observation targets, and at the same time, the invention eliminates the negative influence caused by noise, has good noise robustness in detail, and improves the estimation accuracy of synthetic aperture radar SAR target geometric size.

The invention discloses a 3D terrain imaging system of the interferometric synthetic aperture radar (InSAR) and an elevation mapping method thereof, which mainly solve the problems that the existing InSAR has bad imaging pragmaticality and can not implement 3D elevation mapping on the fast-changing terrain and the transilient terrain. The system comprises three sub-aperture antennas, a radar transmitter, a radar receiver and an imaging data processor; the imaging signal processor comprises a SAR image processing unit and an InSAR image processing unit. The invention receives radar echo throughthe three sub-apertures, then conducts SAR imaging process on the radar echo respectively received by the three sub-apertures, and then conducts InSAR imaging process on the obtained SAR complex pattern, wherein the InSAR imaging process comprises image registration, phase filtering and phase unfolding based on cluster analysis. The processed InSAR phase unfolded image is processed with an elevation inversion to recover a three dimensional digital elevation map. The invention has the advantages of wide adaptability to mapped terrains, and high imaging effectiveness, therefore, the invention can be used in the mapping of the 3D terrain.

The invention provides a channel estimation method and device. Singular value decomposition is performed on an original observation matrix, according to a result of the singular value decomposition, a reconstructed observation matrix is obtained, sparsity is estimated, the sparsity is the number of nonzero elements in a union of a preset channel impulse response value and an index set omega, wherein omega is generated according to the reconstructed observation matrix and a preset residual vector, and finally, channel estimation is performed using the sparsity. Since the singular value decomposition has the characteristic of relatively strong noise robustness, the method and the device provided by the invention have relatively strong noise robustness, in addition, the method and the device can perform estimation of sparsity without the need to know the sparsity, and thus have relatively high practicability, and the channel estimation method and the device provided by the embodiment have relatively high performance.

The invention discloses a feature fusion-based method for suppressing vehicle shadow interference in an open-air scene on a highway. Firstly, the image to be processed is obtained and the foreground region segmentation image is obtained through the background image; secondly, the local region of the foreground region segmentation image is obtained according to the color invariant feature. , and calculate the smoothness of the local area through the gradient information entropy; then calculate the HLGP gradient feature shadow judgment result of the local area according to the local gradient mode direction histogram feature; finally calculate the HLGP of the shadow area according to the HLGP feature shadow judgment result and cosine similarity Shadows interfere with judgment results. The method provided by the invention combines the color grayscale feature and the local gradient coding feature for shadow discrimination and suppression; for the problem of insufficient anti-interference of traditional local features to noise, the influence of noise interference is reduced, and it can adapt to real engineering The application environment, and can obtain more accurate and ideal detection results in this environment.

The present invention proposes a crack image compression sampling method based on a generative adversarial network. The method includes the network architecture design of the generative adversarial network, the crack image generator modeling of the mapping relationship between the crack image and the low-dimensional vector, and the hyperparameters of the adversarial training. Tuning, design of compressed observation matrix for compressed sampling, solution of optimal low-dimensional vector, etc. The method of the present invention adopts the crack image generator trained to generate an adversarial network as a physical constraint to realize the decompression and reconstruction of the image, without requiring the sparsity of the crack image as in the traditional compression sampling method, and has a wider application range. After the generative confrontation network learns the mapping relationship between the crack image and the low-dimensional vector, the low-dimensional vector is optimized based on the gradient descent method, and the fast solution of image decompression and reconstruction is realized. The method has unique advantages in the reconstruction accuracy and reconstruction speed of the fracture image under a relatively high compression rate, and has strong robustness to noise.

The invention discloses a method for identifying the source of a mobile phone in an additive noise environment based on a fusion feature. The fusion feature composed of the MFCC feature, the STFTSDF feature extracted from the Fourier domain and the CQTSDF feature extracted from the CQT domain is used as the device fingerprint , compared with a single feature, it can more accurately characterize the device distinction information; in the training phase, a multi-scene training method is adopted. In the training set, there are not only clean speech samples without scene noise, but also different scene noise types and noise intensities. The M classification model obtained by training is universal, and can effectively identify the source of mobile phones for speech samples of known noise scenes and unknown noise scenes; it uses the deep learning CNN model to establish the M classification model. , the CNN model not only improves the source recognition accuracy of clean speech samples without adding scene noise, but also greatly improves the mobile phone source recognition effect of noisy speech samples, with strong noise robustness.