64results about How to "Improve reconstruction results" patented technology

The invention discloses a method and system of thermal infrared remote sensing image super-resolution reconstruction based on a MAP algorithm, the method and system comprises the following steps: obtaining a segment of sequence thermal infrared band remote sensing images, wherein the sequence images comprise at least two frames of images; completing rectification by using the automatic extraction and matching based on a high precision automatic rectification method of angular point characteristics; achieving the super-resolution reconstruction of the sequence images by using the MAP algorithm, and putting forward an adaptive selection method of an edge penalty function threshold; carrying out application-oriented quality evaluation on target resolution images after reconstruction. According to the method and system of the thermal infrared remote sensing image super-resolution reconstruction based on the MAP algorithm, the high precision automatic rectification among images can be achieved, parameter threshold values can be selected adaptively, the interference of human factors can be reduced, the super-resolution reconstruction of real time can be achieved, and therefore the problems that thermal infrared remote sensing image resolution is low, a reconstruction method cannot achieve automation and is influenced by human factors seriously, speed is not high enough, reconstruction quality cannot be objectively and authentically evaluated in the prior art are solved.

The invention discloses an image super-resolution reconstruction method based on a multi-band deep convolutional neural network, comprising the following steps: selecting training samples and test samples; inputting training images to the network in batches for feature extraction, and combining long-term and short-term memories to construct a multi-band learning structure, then performing featurerefinement, feature mapping and up-sampling reconstruction; and obtaining network parameters according to the model obtained through training to complete the image reconstruction. The invention enhances high-frequency information reconstruction by introducing multi-band learning, and uses the memory migration operation to make the network have long-term and short-term memories at the same time, which speeds up the image reconstruction speed, enhances image edge and texture detail reconstruction, and obtains image super-resolution reconstruction results of better quality. The invention has strong super-resolution capability, and the reconstructed image is closer to the real image.

The invention discloses a method for synchronously reconstructing a dynamic PET image and a tracer kinetic parameter on the basis of TV (total variation) and sparse constraint. The method introduces a dictionary based on sparse constraint and a TV operator to carry out a whole reconstruction process by establishing a mathematic model for synchronously estimating the dynamic PET image and the kinetic parameter, wherein a dynamic PET sequence reconstruction subproblem provided with the total variation operator is subjected to iterative optimization solution by an ADMM algorithm, and a PET kinetic parameter estimation subproblem in combination with the dictionary based on sparse constraint is solved by a soft threshold iterative algorithm. The method solves a problem for simultaneously estimating the dynamic PET image sequence and the kinetic parameter, introduces the TV operator to improve the low result resolution and noise interference proneness in the PET image reconstruction process, and may obtain a better reconstruction result compared with other algorithms which individually reconstructs the PET image or estimates the kinetic parameter.

The invention discloses a thermal analysis test system based on the infrared sequence image super-resolution reconstruction method. A photographing device, which is an non-refrigerating thermal infrared imager, is fixed on the x-y displacement platform of a photographing frame; a target which is in conducted state, such as a circuit, is fixed on the supporting pillar of the photographing frame. The non-refrigerating thermal infrared imager shoots the target to get the infrared sequence image of the target which is stored in a computer terminal through an interface arranged on the non-refrigerating thermal infrared imager. The super-resolution reconstruction method provided by the invention can be used to conduct the super-resolution reconstruction to the infrared sequence image of the target to get the resolution image of the target for analyzing. Under the condition that the photographing distance is fixed, the thermal analysis test system can get images, the resolution of which is much higher than the physical resolution of the photographing device.

Machine learning techniques are used to train a “dictionary” of input signal elements, such that input signals can be linearly decomposed into a few, sparse elements. This prior knowledge on the sparsity of the input signal leads to excellent reconstruction results via maximum-aposteriori estimation. The machine learning imposes certain properties on the learned dictionary (specifically, low coherence with the system response), which properties are important for reliable reconstruction.

The invention discloses a joint reconstruction method for multiple dynamic contrast enhancement nuclear magnetic resonance images. According to the method, a conjugate gradient algorithm is combined with a variable density random sampling method to reconstruct the DCE-MRIs in a high probability and efficiency mode from local k-space sampling data. The method includes the following steps of Fourier transform of the images, design of a sampling template, downsampling through the sampling template, inverse Fourier transform, design of wavelet sparse transformation matrix coefficients, design of a constrained energy function, conversion of the constrained energy function into a unconstrained problem through a Laplace operator, acquisition of a solution through the conjugate gradient algorithm based on QUOTE l2,1 l2,1 norms, quantitative evaluation of an obtained reconstruction result of the DCE-MRIs, and difference value contrast of the reconstruction result and the organic images. The reconstruction speed of the images is faster, accuracy and definition are higher, and according to the result, a good contrast mid-value sequence, the high signal to noise ratio, sufficient analysis coverage and rapid data acquisition are achieved.

The invention discloses a dynamic PET image reconstruction method based on self-encoder image fusion. The method learns from the theory of integrated learning in machine learning, takes an MLEM algorithm as a weak classifier, obtains a strong classifier through integration of different weak classifiers, and improves the PET reconstruction effect. The method improves an existing MLEM algorithm, adopts a self-encoder structure to carry out image fusion on reconstruction results of different iterations, so that obtains an optimized reconstruction result. Compared with reconstruction methods in the prior art, the method has better reconstruction effect.

The present invention discloses a method for realizing three-dimensional reconstruction by multi-spectral image fusion, comprising the following steps of: realizing biorthogonal wavelet decomposition; determining fusion operators; implementing HIS transform, and separation of space and spectral characteristics; combining bands; fusing a resulted image. After processing with the method for realizing three-dimensional reconstruction by multi-spectral image fusion, an internal structure of a three-dimensional structure is quite clear, and therefore digital effect of the multi-spectral image is optimal, and complementation of advantages of single spectrum information is realized.

The present invention relates to the novel template image super-resolution technology. The method comprises: obtaining an original image set, performing 0.5-times downsampling of each image in the image set, and obtaining low-resolution images; performing outline template interpolation amplification for two times, and obtaining high-resolution reference images; performing novel template enhancement of the high-resolution reference images and the low-resolution images; performing matching of SIFT feature points, and calculating a motion matrix between the high-resolution reference images and the low-resolution images so as to obtain corresponding points in the low-resolution images of the points on the high-resolution reference images; and calculating the residual errors of the points on the high-resolution reference images and the corresponding in the low-resolution images to combine the residual images, performing overlapping of the residual images and the high-resolution reference images, and obtaining final images. The novel template can effectively enhance image high-frequency information and effective maintain low-frequency information in the images so as to have wide application in the image enhancement field.

The invention discloses a Fourier laminated microscope pupil recovery method based on a neural network in the field of computer imaging. The problem that the reconstruction precision of the existingFourier laminated imaging model is low under the influence of optical phase difference is solved. A neural network model is established based on a TensorFlow deep learning framework in combination with a forward imaging mode of an FPM system. The problem that the universality of a reconstruction model based on a deep convolutional neural network is poor is solved; a recovery process for a pupil function of the system is introduced, so that the influence of optical aberration in the system on a reconstruction result can be better suppressed, and a better result is obtained. According to the invention, the frequency spectrum and the pupil function of the sample are set as a trainable two-dimensional network layer in the network; complex amplitude information and a pupil function of a sampleare obtained at the same time by minimizing a loss function in the training process. The method has good universality and can still obtain a reconstruction result better than that of a traditional algorithm under the condition that aberration exists in the system.

The invention discloses a single-frame image super-resolution reconstruction method based on a natural image statistic sparse model. The technical problem that the quality of reconstruction images of an existing single-frame image super-resolution reconstruction algorithm based on sparse representation is poor is solved. According to the technical scheme, by the aid of statistical properties of a natural image, the modeling is performed for image super-resolution reconstruction through a Bayes method, and a high-resolution image is estimated through a minimum mean-squared error standard. The obtained high-resolution image through reconstruction is natural, the pseudo-structure number is reduced, and a clear edge structure is achieved. Compared with a background technology method, the method has the advantages that super-resolution reconstruction images with high quality are obtained, and reconstruction results are improved by 1 dB to 2 dB.

The invention discloses a super-resolution image reconstruction method and system based on multi-feature learning, and the method makes full use of the rich information contained in a single input image for reconstruction, and does not depend on an external database. According to the method, a mapping relation between image features is established based on cross-scale similarity of images, and a high-resolution image containing high-frequency information is reconstructed for an input image directly by using the mapping relation, so that the defect of high-frequency information loss caused by image reconstruction by using an interpolation amplification method is well overcome. According to the method, effective high-frequency information is acquired by using singular value thresholding, andthe high-frequency information is amplified by using a gradient feature mapping relation and then is overlapped on a high-resolution image in a blocking manner, so that a final image reconstruction result is obtained. According to the method for reconstructing the image by utilizing the image feature combination, noise points of the reconstructed image are effectively inhibited, image edge and texture information is well kept, and detail enhancement of the image is realized.

The invention discloses an interactive three-dimensional reconstruction method based on an image, and the method comprises the steps that a camera obtains a plurality of frames of images, and transmits the plurality of frames of images to a processor; a processor judges the received multiple frames of images and screens out multiple key frames from the multiple frames of images; the processor performs incremental analysis on the plurality of screened key frames and performs closed-loop detection at the same time; the processor is used for carrying out model consistency judgment on incrementalanalysis and closed-loop detection results and giving an interaction prompt when models are inconsistent; the processor provides a plurality of interactive interfaces (3D-3D alignment, 3D-2D alignmentand 2D-2D matching), so that a user can input key information conveniently; the processor fuses key information input by a user with incremental analysis and closed-loop detection results to completethree-dimensional reconstruction. According to the method, an automatic processing algorithm is properly combined with user interaction, and the accuracy of a reconstruction result is improved.

The invention discloses a simple image reconstruction method based on LFP phase features and a K-nearest neighbor algorithm and relates to the field of information processing technology. The method comprises the steps that local field potential signals of a brain visual cortex under image stimulation are collected, and sample stimulation data, target stimulation data, sample response data and target response data are obtained after the image stimulation and the local field potential signals are processed respectively; according to the sample response data and the target response data, sample phase features and target phase features are acquired, and a sample response matrix and a target response matrix are constructed; an image decoder is constructed according to the sample stimulation data and the sample response matrix through the K-nearest neighbor algorithm; and the target response matrix is substituted into the decoder, and a simple image is obtained after target decoding stimulation data is acquired. Through the method, the problems that in existing image stimulation reconstruction based on biological vision, the image reconstruction experiment process is complicated, and a reconstructed image has a poor effect and is not clear are solved.

The invention discloses a low-dose CT reconstruction method based on a twin feedback network, and belongs to the field of medical image processing and computer vision. According to the invention, themethod comprises the steps: taking a low-dose CT image as input, acquiring a preliminary reconstruction result through a reconstruction branch of the twin feedback network, acquiring a boundary graphthrough a priori branch, continuously optimizing the network parameters through detail protection loss, and obtaining a final reconstruction result. The method is simple in program and easy to implement, and a high-resolution reconstruction result can be obtained in an end-to-end mode. The detail protection loss function designed by the invention can comprehensively realize CT image detail protection from two aspects of local and global aspects.

The invention discloses a hyperspectral image super-resolution method based on a deep closed-loop neural network. The method comprises two processes, namely constructing a deep closed-loop neural network model for hyperspectral data, and reconstructing a high-resolution hyperspectral image through variable separation and fine optimization. Two deep learning models are constructed to learn a super-resolution process and an inverse super-resolution process respectively, and form a closed-loop network to reduce a mapping space, and promoting model fitting; a network structure suitable for a hyperspectral image is adopted to extract spatial features and spectral features, spatial information and spectral information are jointly reconstructed, and the quality of the image obtained through super-resolution is improved; and model iterative solution is performed by using the trained deep closed neural network and adopting a variable separation fine optimization method, and a reconstruction result is optimized. According to the method, the deep closed-loop neural network suitable for hyperspectral image super-resolution is used, the mapping space can be reduced through the closed-loop network, and a reconstruction result better than that of a one-way network can be obtained.

The invention discloses a three-dimensional object reconstruction algorithm based on deep learning, and the algorithm comprises the steps: inputting a plurality of object two-dimensional images obtained from any angle, carrying out the preprocessing, building a convolutional neural network, enabling the two-dimensional images to serve as training data, inputting the training data into the built convolutional neural network for training, and inputting a to-be-measured two-dimensional image into the trained convolutional neural network model, and outputting a three-dimensional reconstruction result by the convolutional neural network model. According to the invention, the convolutional neural network model comprises an encoder, a decoder and a multi-view feature combination module. The input of the encoder is a multi-view two-dimensional image, the output of the encoder is a two-dimensional feature vector, and the two-dimensional feature vector needs to be converted into three-dimensional information; the three-dimensional information is input into a decoder to obtain three-dimensional prediction voxel occupation of the single image; and finally the final predicted voxel occupation is obtained through a multi-view feature combination module. In the test stage, the accuracy is calculated according to the 0-1 occupancy predicted by the hierarchical prediction strategy and the real ground occupancy.

The invention discloses a three-dimensional medical image super-resolution reconstruction method and system, and belongs to the field of medical image processing. According to the method, a three-dimensional super-resolution problem is decomposed into combination of super-resolution reconstruction of a single slice and content correlation between adjacent slices, super-resolution reconstruction is carried out on three-dimensional medical image data through the multi-channel two-dimensional convolutional neural network, and high correlation between the adjacent slices is fully considered in a multi-channel network structure. Parameters of an image super-resolution reconstruction part are trained through a large number of two-dimensional high-resolution medical images, then the parameters of the reconstruction part are frozen, and a small amount of three-dimensional data is used for training weight parameters of a multi-channel output layer. Because the multi-channel network only needs to train parameters such as the weight of the sequence image of the new output layer, network training can be completed only through a small number of three-dimensional high-resolution images. The two-dimensional network is adopted essentially, compared with a three-dimensional network, the required three-dimensional data is greatly reduced, and the training difficulty is also greatly reduced.

The invention relates to a method for reconstructing a three-dimensional human body with shielding based on depth map restoration, which is used for solving the problem of restoring the posture, the body shape and the surface details of a three-dimensional human body model from a single RGB human body image containing a shielding object. The method comprises the following steps: a basic model construction stage: constructing a basic three-dimensional human body model through an existing SMPL model construction method; a body shape posture optimization stage: optimizing three-dimensional human body joint point positions through a Joint encoder network, and optimizing three-dimensional human body shape characteristics through an Anchor encoder network; a surface detail optimization stage: providing a DHDNet network structure, and recovering three-dimensional human body surface details by reconstructing a complete human body depth image. Meanwhile, a DepthHuman data set containing a large number of human body RGB images and synthesized human body depth images is constructed. The effect of reconstructing the complete three-dimensional human body model containing rich details only through the single RGB human body image containing the shielding object is achieved.