53results about How to "Improve rebuild quality" patented technology

The invention discloses a large-scale MIMO channel state information feedback method based on deep learning. The method comprises the following steps: firstly, carrying out two-dimensional discrete Fourier transform (DFT) on a channel matrix H-wave of MIMO channel state information in a spatial frequency domain on a user side, so that a channel matrix H which is sparse in an angle delay domain isobtained; secondly, constructing a model CsiNet comprising a coder and a decoder, wherein the coder belongs to the user side and is used for coding the channel matrix H into codons with a lower dimension, and the decoder belongs to a base station side and is used for reconstructing an original channel matrix estimation value H-arrow from the codons; thirdly, training the model CsiNet to obtain model parameters; fourthly, carrying out two-dimensional inverse DFT on a reconstructed channel matrix H-arrow which is output by the CsiNet, so that a reconstructed value of the original channel matrixH-wave in the spatial frequency domain is recovered; and finally, using the trained model CsiNet for compressed sensing and reconstruction of channel information. The method provided by the inventionhas the advantages that large-scale MIMO channel state information feedback expenditures can be reduced, and an extremely high channel reconstruction quality and an extremely high channel reconstruction speed can be achieved.

The invention discloses a reconstruction method for image super-resolution. The reconstruction method comprises the following steps of: (1) inputting a low-resolution image sequence of m frames represented by vectors in the same scene; (2) processing all low-resolution images by adopting a bilinear interpolation method, and acquiring initial high-resolution image estimated values, namely initial estimated values; (3) substituting the initial estimated values into an iterative formula, and performing iterative computation; (4) judging whether iterative convergence conditions are met or not, stopping iteration if the iterative convergence conditions are met, and outputting a high-resolution image. The reconstruction method for the image super-resolution provided by the invention has wide application prospects in the fields such as video and image processing, medical imaging, remote sensing images and the like.

The invention discloses a partial echo compressed sensing-based quick magnetic resonance imaging (MRI) method. The conventional imaging method has low speed and high hardware cost. The method comprises the following steps of: acquiring echo data of a random variable density part, namely intensively acquiring data in a central area of a k-space and acquiring the data around the k-space randomly and sparsely to generate a two-dimensional random mask, adding the two-dimensional random mask into every data point which needs to be acquired on a frequency coding shaft to form a three-dimensional random mask, and acquiring the data of the k-space according to the generated three-dimensional random mask; re-establishing by projection onto convex sets based on a wavelet domain which is de-noised by soft thresholding; and nonlinearly re-establishing a minimum L1 normal number based on finite difference transformation, namely sparsely transforming an image space signal x, determining an optimization objective and solving the optimization objective. By the method of the invention, partial echo technology and compressed sensing technology are combined and applied to data acquisition of MRI, sothat echo time is shortened, and data acquisition time is shortened at the same time.

The invention discloses a method of improving the density of the three-dimensional reconstructed point cloud based on neighborhood block matching. The method comprises the steps of obtaining the rough and sparse object point cloud using a three-dimensional reconstruction algorithm based on an image sequence, and obtaining the transformation matrix of a camera in the three-dimensional space when each frame of images is taken; processing the original image again, and performing the dense feature matching in the image using a neighborhood-based block matching algorithm; then, according to the obtained position of the camera in the space, carrying out a legitimacy test on the obtained dense feature points, and mapping the feature points satisfying requirements to corresponding positions in the three-dimensional point cloud; and conducting an outer point filtering for the obtained point cloud using an outer point deletion algorithm based on the object contour, and performing a color remapping to obtain the dense point cloud which is far better than the original point cloud. According to the invention, the point cloud having far higher quality and density than that by the traditional algorithm can be obtained, the effect of the original algorithm can be greatly improved, and the quality of reconstruction is improved. The method has high universality and robustness.

The invention relates to a cross-node controlled online video stream selective retransmission method comprising the following steps of: when an online-coded single video stream enters a transmitting end, dividing the online-coded single video stream into DFS (Delay Constraint Frame Sets) one by one according to starting and transmitting delay limitation as selective retransmission basic objects; using the transmitting end as a first communication node, determining importance levels of IP (Internet Protocol) packets in the DFS according to fault-tolerant importance measure, and supplying the selective retransmission of the IP packets by packet scheduling based on fault-tolerant importance; and on the basis of the importance levels of the IP packets, using a base station as a second communication node, and supplying the selective retransmission of a wireless link unit by a priority descending ARQ (Automatic Request for Repetition) mode. Under delay and bandwidth condition limitation, the invention organically combines priority transmission mechanisms among main communication nodes, realizes the importance division and the selective retransmission of wireless link unit levels, can strengthen the error control performance of online video streams and effectively increase the rebuilding quality of received videos.

Distributed compressed sensing coding of video signals is the base and the key for realizing high-efficiency transmission of the video signals. Side information is acquired by motion search and homographic matrix compensation time interpolation by aiming at the problems of side information accurate acquisition and reconstruction algorithm in distributed compressed sensing coding, and decoding can be performed through fusion of the side information after the coset index is received by a decoding end. Smooth constraining is performed on decoding pixel points by introducing a Huber-Markov random field prior probability model in joint decoding reconstruction so as to obtain better reconstruction quality. According to a distributed video coding and decoding technical method based on compressed sensing, the method of "independent coding and joint decoding" is adopted, and complexity is transferred to the decoding end from a coding end so that complexity of coding operation is enabled to be low, and the method is suitable for resource constrained video coding equipment in a wireless network.

The invention relates to a novel viewpoint rebuilding method in a multi-viewpoint collection/display system of a converge type camera. According to a collection mode of the converge type camera, first, a geometric model of the converge type camera is built, each parameter related to the collection system is comprehensively considered, an optimized condition of distortionless rebuilding viewpoints is obtained through signal sample theory analysis, then spectrum analysis is performed on EPI image on the scene, relation between scene EPI spectral characteristic and scene depth span is obtained. Based on the above analysis, the number of the sample image is determined, and a viewpoint rebuild filter is designed, a sample image is generated for the rebuilt viewpoint through interpolation technology, finally a novel viewpoint of a three-dimension scene in certain range is obtained through the converge model designing and plotting rebuild method. In experiment, verification is respectively performed on an analogue system and a practical system with similar parameters, obtaining excellent rebuilding quality. The invention is of reference value for other camera array types and viewpoint rebuilding system.

The invention discloses a face video compression method and belongs to the technical field of face video data compression in the case of multimedia communication. The method is characterized by comprising steps: a face model is used for carrying out positioning and parametric characterization on the face in the video at the sending end, a face model parameter vector represented by an illumination parameter vector, a gesture parameter vector and a shape and appearance combined parameter vector is obtained, constraint that global minimal parameter error and model parameter total bit are less than the given upper limit on the basis of the given to-be-retained parameter frame sample number is carried out, the face model parameter vector after optimal adaptive interframe compression is sent to a receiving end, an independent slicing linear interpolation method is used for restoring the face model parameter vector, and the original image is obtained through face shape calculation and restored face appearance. The face video compression ratio is greatly improved while the optimal reconstruction quality is ensured, redundancy in the face video is removed to the maximal degree in the time domain, and the communication resource utilization rate is thus improved.

The invention discloses a complex amplitude hologram algorithm based on a light field, and the algorithm comprises the steps: firstly carrying out the multi-angle photographing of a three-dimensionalobject through a virtual camera, and obtaining a series of subgraphs; then adding an inclination phase factor of a corresponding angle to each sub-image in an algorithm, and adding the obtained imagesto obtain a complex amplitude hologram; or, displacing the frequency spectrum of each sub-image in the frequency domain space, enabling the displacement to correspond to the angle information duringshooting, and adding the sub-images of which the frequency spectrum positions are moved to obtain a complex amplitude hologram; and finally, modulating the obtained complex amplitude hologram by usinga complex amplitude modulation device so as to reconstruct the three-dimensional object. According to the algorithm, a three-dimensional light field closer to a natural scene can be reconstructed, the calculation speed of a hologram is increased, and meanwhile, a reconstruction light path is simplified.

An embodiment of the invention provides an ultrasonic three-dimensional reconstruction method and device. The method comprises: determining dimensions of reconstruction volume data, wherein the dimensions include a coordinate origin of the volume data, dimensional size of the volume data and inter-voxel physical spacing; traversing each pixel in a two-dimensional ultrasonic plane corresponding to a two-dimensional ultrasonic image sequence, mapping each pixel of the two-dimensional ultrasonic plane into three-dimensional volume data through Bessel interpolation algorithm, and carrying out interpolation; performing hole completion on the three-dimensional volume data through a kernel regression algorithm. The ultrasonic three-dimensional reconstruction method and device provided by the embodiment of the invention enable reconstruction quality of three-dimensional reconstruction to be improved effectively by the combination of the Bessel interpolation algorithm and kernel regression algorithm.

The invention discloses a satellite cloud picture super-resolution reconstruction method based on deep learning, and the method comprises: making and preprocessing a satellite cloud picture, and obtaining a high-resolution satellite cloud picture data set; dividing the high-resolution satellite cloud picture data set into a training data set and a test data set for modeling; constructing and training a satellite cloud picture super-resolution reconstruction model based on deep learning; and inputting the low-resolution satellite cloud picture into the trained satellite cloud picture super-resolution reconstruction model to obtain a high-resolution satellite cloud picture. The high-precision detailed satellite cloud picture high-resolution reconstruction image is obtained through the deep learning method, the reconstruction effect which is more accurate and universal than that of a traditional method is obtained, the high-resolution satellite cloud picture can be obtained through the low-resolution satellite cloud picture, and the practicability of the method is improved.

The invention relates to a distributed multi-view video compression sampling reconstruction method combining space-time side information, and aims to meet the application requirements of low-complexity video acquisition coding under the condition that the computing and storage capacities of all nodes in a distributed multi-view video acquisition network are limited. Considering that continuous video frames in the same viewpoint have time correlation, the video frames of the adjacent viewpoints at the same moment have spatial correlation, and the time and space side information obtained by thetime correlation and the spatial correlation can be further characterized by sparse constraints of differences between the side information and the current frame, so that a non-key frame reconstruction optimization model in the multi-viewpoint video is generated. And finally, optimization solution is carried out through an FISTA strategy, so that non-key frame information with relatively good reconstruction quality is generated.

The invention discloses a channel state information feedback method based on deep learning and entropy coding, and the method comprises the steps: carrying out the preprocessing of a channel matrix of MIMO channel state information at a user side, selecting key matrix elements to reduce the calculation amount, and obtaining a channel matrix H which is actually used for feedback; constructing a model combining a deep learning feature encoder and entropy encoding at a user side, and encoding a channel matrix H into a binary bit stream; at a base station end, constructing a model combining a deep learning feature decoder and entropy decoding, and reconstructing an original channel matrix estimated value from a binary bit stream; training the model to obtain model parameters and an output reconstruction value of a reconstructed channel matrix, and finally applying the trained model based on deep learning and entropy coding to compressed sensing and reconstruction of channel information. According to the invention, the large-scale MIMO channel state information feedback overhead can be reduced.

The invention relates to a remote sensing image super-resolution method based on a multi-scale feature adaptive fusion network, and the method comprises: 1) carrying out the convolution operation of an originally inputted low-resolution remote sensing image through a filter, and extracting an original feature map; 2) extracting self-adaptive multi-scale features of the original feature map throughn cascaded multi-scale feature extraction modules AMFE to obtain a self-adaptive multi-scale feature map; 3) superposing the original feature map and the adaptive multi-scale feature map, and performing convolution operation on the superposed map by using a filter to realize feature dimension reduction and fusion; and 4) by adopting a sub-pixel convolution method, obtaining a final remote sensingimage after super-resolution reconstruction. The invention provides the remote sensing image super-resolution method based on the multi-scale feature self-adaptive fusion network, which can realize self-adaptive fusion of multi-scale feature information of the remote sensing image, can realize efficient reconstruction of high-resolution detail information of the remote sensing image and can improve the image super-resolution reconstruction effect.

The invention discloses a compressed sensing image and video recovery method based on tensor approximation and space-time correlation, and relates to the technical field of video image processing. Themethod comprises the following steps: S1, material reading: reading a material, and carrying out framing processing to form an independent image; S2, image blocking: performing image blocking processing on the single image in the step S1 to obtain small image blocks; S3, recovery processing: carrying out initial image recovery processing on the small image blocks by adopting a tensor approximation thought; and S4: image reconstruction: carrying out image reconstruction processing on the image after initial image recovery processing to obtain an image block with good quality. According to themethod, a novel recovery model is provided for natural images and videos, and high-quality recovery is realized; and according to the image restoration model, local self-similarity can be fully utilized by extracting similar blocks, and the model can be established by considering low-rank attributes, so that the quality of a reconstructed image is further improved.

The invention provides a face image compression method and device based on deep learning, and relates to the technical field of image processing, and the method comprises the steps: decomposing a testface image into a shape component and a texture component based on an average face image and an active appearance model; then, using a quantization entropy encoder for encoding and compressing the shape component to obtain a first compression result; using a trained convolutional neural network encoder for encoding and compressing the texture component to obtain a second compression result; and finally, determining the first compression result and the second compression result as a face image compression result. According to the method, the average face image and the active appearance model are used as priori knowledge, and the trained convolutional neural network encoder is used to encode and compress the texture component, so that the second compression result represented by a low-dimensional feature can be obtained, therefore, the redundancy of the texture component is reduced, and the compression efficiency of the face image is improved.

The invention provides a multiple description coding high-quality edge reconstruction method based on spatial downsampling, which comprises the following steps: making a data set: selecting a video, dividing the video into two descriptions through spatial downsampling, coding and decoding the video under the setting of a quantization parameter QP value, and taking the decoded video and a corresponding original video as a training set; and training a SD-VSRnet network: taking every five frames of the video as an input of the network, sequentially performing feature extraction, high-frequency detail recovery and pixel rearrangement, then performing jump connection with an input intermediate frame to obtain reconstructed video frames, and performing frame-by-frame reconstruction to obtain a final reconstructed video, thereby realizing training of the SD-VSRnet network. According to the method provided by the invention, a multiple description coding high-quality edge reconstruction data set suitable for space downsampling is made, in addition, a video super-resolution neural network is adopted to separately test four QP values, and the edge decoding video reconstruction quality of different compression degrees can be effectively improved.

The invention discloses a measuring device for measuring gas parameters of a combustion field, belongs to the technical field of flow field optical measurement, and can solve the problem of poor measurement result caused by insufficient light distribution in the flow field in the prior art. The measuring device comprises a measuring ring; each laser emitting unit corresponds to each laser receiving unit group, and the corresponding laser emitting unit and laser receiving unit group are respectively arranged on two opposite side walls of the measuring ring; each laser receiving unit group comprises a plurality of laser receiving units arranged in rows; each single-mode optical fiber corresponds to each laser emission unit; each first multimode optical fiber corresponds to each laser receiving unit; the single-mode optical fiber outputs a laser beam to the laser emission unit; the laser emitting unit converts the laser beam into a fan-shaped light beam, and the laser receiving unit converges the fan-shaped light beam irradiated on the laser receiving unit and couples the fan-shaped light beam into the corresponding first multimode optical fiber. The device is used for flow field gasparameter measurement.

The invention discloses a panoramic video coding optimization algorithm based on a user field of view, and the algorithm comprises the following steps: (1), obtaining a predicted FoV mode FoV([mu]i, [mu]j); (2) generating a two-dimensional Gaussian distribution probability corresponding to the FoV (i, j) in the current frame according to the FoV([mu]i, [mu]j); (3) obtaining the cumulative probability that the tile (i, j) is possibly in n*m FoV modes, namely the probability that the tile is observed; (4) calculating the projection weight of tile (i, j) by using a projection transformation formula in an ERP format; (5) calculating and correcting lambda and QP values according to the finally obtained observation probability and projection weight; and (6) encoding by using the finally obtained QP value, and encoding according to the obtained tile-level quantization parameter selection scheme after encoding. According to the method, the coding quality of a user watching area is fully considered, so that the final coding effect achieves better reconstruction quality under the condition of consuming smaller bits.

The invention discloses an extensible man-machine cooperation image coding method and coding system. The method comprises the following steps: extracting an edge graph of each sample picture and vectorizing the edge graph as a compact representation for driving a machine vision task; performing key point extraction in the vectorized edge image to serve as auxiliary information; respectively carrying out entropy coding lossless compression on the compact representation and the auxiliary information to obtain two paths of code streams; carrying out preliminary decoding on the two paths of code streams to obtain an edge graph and auxiliary information; inputting the edge graph obtained by decoding and auxiliary information into a generative neural network, and performing forward calculation of the network; carrying out loss function calculation according to the obtained calculation result and the corresponding original picture, and carrying out back propagation on the calculated loss to a neural network to carry out network weight updating until the neural network converges, so as to obtain a double-path code stream decoder; obtaining an edge image and auxiliary information of a to-be-processed image, and encoding and compressing the edge image and the auxiliary information to obtain two paths of code streams; and enabling the two-path code stream decoder to decode the received code stream and reconstructs an image.