50results about How to "Improve the effect of super resolution" patented technology

The invention discloses a super-resolution image generation method. The super-resolution image generation method includes inputting a real image sample into a generation network so as to output a super-resolution image sample after the generation network and a distinguishing network are preset, acquiring distinguishing probabilities outputted by the distinguishing network after the real image sample and the super-resolution image sample are outputted, determining generation network loss functions and distinguishing network loss functions according to the real image sample, the super-resolution image sample and the distinguishing probabilities, and adjusting configuration parameters of the generation network and the distinguishing network according to the generation network loss functions and the distinguishing network loss functions; receiving a low-resolution image to be processed after adjustment is completed, generating a super-resolution image of the low-resolution image according to the generation network, and subjecting the super-resolution image to visualized processing. By the super-resolution image generation method, image super-resolution effect and realness of the super-resolution image are both improved remarkably.

The invention relates to an image super-resolution method based on a channel attention mechanism and multilayer feature fusion, and the method comprises the steps of directly extracting the original features of a low-resolution image at the beginning of a residual branch by using a single-layer convolutional layer based on deep learning; using six cascaded convolutional circulation units based ona channel attention mechanism and multi-layer feature fusion to extract accurate depth features; carrying out upsampling on the depth features through a deconvolution layer, and carrying out dimensionality reduction on the upsampled features through a single-layer convolution layer to obtain a residual error of the high-resolution image; carrying out up-sampling on the low-resolution image by using a bicubic interpolation method in a mapping branch to obtain mapping of the high-resolution image; and adding the mapping and the residual of the high-resolution image pixel by pixel to obtain a final high-resolution image. The method is reasonable in design, fully considers the difference between the feature channels, efficiently utilizes the hierarchical features, and maintains a higher operation speed while obtaining higher accuracy.

The invention discloses a video live broadcast method for wearable devices, comprising processing at a sender and processing at a receiver. The sender decreases the spatial resolution of a high-definition video signal four times and then executes standard video encoding so as to greatly reduce the pressure put on a wireless network by the encoding data, and the receiver restores the proper resolution through super-resolution reconstruction. Time-domain motion matching needed for super-resolution reconstruction is speeded up with the help of inherent motion vector information of a video encoder. Moreover, a reconstruction residual error compensation link is introduced into a video codec loop to eliminate distortion between a super-resolution high-definition image and a real high-definition image. The mobile video live broadcast efficiency of wearable devices is improved greatly under the premise of no additional damage to the quality of high-definition videos.

The invention relates to a fine-grained scale image super-resolution method based on a non-local enhancement network, and the method comprises the following steps: A, carrying out the preprocessing ofan original high-resolution training image, and obtaining an image block pair data set composed of low-quality high-resolution image blocks of different scales and original high-resolution training image blocks; B, training a non-local enhanced deep network for the data set by using the image blocks; C, inputting the high-resolution image of the low-quality test image into a deep network for reconstruction to obtain a super-resolution result. According to the method, a non-local enhanced deep residual structure is used, non-local operation and common convolution are combined, local and non-local image attributes can be effectively captured, image super-resolution is carried out, and compared with an existing super-resolution model, the method can remarkably improve the performance of image super-resolution on the fine-grained scale.

The invention discloses a single image super-resolution method based on a reversible network, and belongs to the field of image processing. According to the method, a network structure of a super-resolution model is constructed by introducing a reversible network; mutual mapping of a high-resolution image space and a low-resolution image space is realized by utilizing the reversible property of areversible network; the super-resolution process is optimized in the low-resolution direction and the high-resolution direction, the problem that other super-resolution methods based on deep learningcannot effectively utilize the mutual dependence between high-resolution images and low-resolution images is solved, and therefore the image super-resolution capability of the model is improved. A weight matrix of the 1 * 1 reversible convolution layer is initialized by introducing singular value decomposition, so that the propagation speed of the inverse process of the 1 * 1 reversible convolution layer is increased; by the adoption of the method, the super-resolution process of a single image can be effectively achieved, and the super-resolution image with good texture details and visual effects is generated through the low-resolution image.

The invention discloses an image super-resolution method based on a densely linked neural network, a storage medium and a terminal. The method includes: preprocessing an image; performing Feature extraction: Building a dense-linked neural network, inputting the low-resolution image Input from the entrance of the dense-linked neural network, and extracting the feature information contained in the Input after calculation; Predicting the super-resolution image and updating the network parameters: performing upsampling / deconvolution on the feature-extracted image to obtain the predicted image predict; calculating the error values between the predicted image predict and the real image label, and updating the parameters of the densely linked neural network in the reverse direction; and performing super resolution reconstruction. The method can remarkably improve the ability of extracting the low-frequency and high-frequency features of an image by a depth neural network, improve the effect of the image super-resolution, and improve the ability of providing information by a picture, so that the invention is applied in the field of expecting to obtain a high-resolution image and providingmore details by the picture.

The invention discloses a multi-frame video super-resolution method fused with an attention mechanism. The method comprises the following steps: collecting video data and training the video data by adopting a video enhancement technology to generate a training set and a test set; connecting the deformation convolution feature alignment module and the feature reconstruction module to form a multi-frame super-resolution network, and training the multi-frame super-resolution network by using the training set; adding a 3D convolution feature alignment module into the multi-frame super-resolution network, and training the multi-frame super-resolution network by adopting the training set; adding the feature fusion module into a multi-frame super-resolution network, and training the multi-frame super-resolution network by adopting the training set; finely adjusting the multi-frame super-resolution network by adopting the training set to generate a multi-frame super-resolution model; and testing the multi-frame super-resolution model by adopting the test set. According to the invention, the super-resolution effect can be effectively improved by analyzing big data.

The invention discloses an image restoration method based on an enhanced neural network. The method comprises the following steps: S1, converting an image to be restored into a plurality of low-resolution images under different zoom factors; S2, converting the image to be restored into a plurality of low-resolution images under different zoom factors; S3, converting the image to be restored into aplurality of low-resolution images. S2, respectively inputting a plurality of low-resolution images to a first depth convolution neural network trained in advance, thereby obtaining a plurality of high-resolution images under corresponding different scaling factors; S3, converting a plurality of high-resolution images in S2 into images having the same size as the image to be restored, and fusingthese images to obtain the restored images. The invention also discloses a corresponding storage medium and an image restoration system. The invention can prevent the network from being degraded in the training process and accelerate the convergence speed.

The invention discloses a single-frame image super-resolution reconstruction method based on cascade regression base learning. The method comprises the following steps: taking a super-resolution reconstruction technology of a single-frame low-resolution image as a research object, learning a multi-layer over-complete sub-dictionary for representing an image structure, constructing a mapping relation between a low-resolution image and a high-resolution image, and learning an optimized regression base and a corresponding coding coefficient; and then, complete super-resolution reconstruction is realized for the low-resolution image set, and the reconstructed image is used as a low-resolution image of the next layer for feature extraction. The invention discloses a single-frame image super-resolution reconstruction method. learning by utilizing a meta-dictionary learning method to obtain a low-resolution dictionary; a weighted linear regression method is used for carrying out multilayer regression base learning on a reconstructed high-resolution training set image and an original high-resolution image in a cascading mode so as to approach a complex nonlinear mapping relation between alow-resolution image and a high-resolution image, and instance regression super-resolution reconstruction with high processing speed, small memory occupation and high reconstruction quality is achieved.

The invention discloses an image super-resolution method based on feature correlation, a storage medium and terminal equipment, the image super-resolution method applies a preset super-resolution network, and the super-resolution network comprises a feature extraction module, a residual block attention module and an image reconstruction module; the method comprises the following steps: inputting alow-resolution image into a feature extraction module, and outputting a shallow feature map through the feature extraction module; inputting the shallow feature map into a residual block attention module, and performing spatial and channel feature correlation learning through the residual block attention module to obtain a depth feature map; and inputting the depth feature map into an image reconstruction module, and outputting a high-resolution image through the image reconstruction module. According to the invention, spatial and channel feature correlation learning is carried out on the feature image through the residual block attention module, and the expression capability of the feature image is enhanced, so that the super-resolution effect of the image is improved.

The invention relates to the technical field of image processing, and provides an image super-resolution method and device, terminal equipment and a computer storage medium. The method comprises the following steps: extracting image features from a to-be-processed image, inputting the image features into a texture feature extraction module to extract texture features of the image, processing the texture features by adopting a plurality of convolution layers to obtain structural features of the image, fusing the texture features and the structural features of the image, and finally recovering the fused texture features and structural features into an up-sampled image; and obtaining a corresponding super-resolution image. Compared with a traditional image super-resolution method, the image super-resolution method has the advantages that richer image details can be obtained by fusing the texture features and the structural features of the image, and therefore the image super-resolution effect is improved.

The invention provides an SAR image super-resolution method based on a neural network, and belongs to the field of image super-resolution. Aiming at a set neural network structure, a loss function isconstructed by utilizing a mean square error between a predicted value and a real value, a model mapping problem is converted into an optimization problem of the loss function, weights and bias valuesof all layers of the neural network are determined, and finally low resolution-high-resolution mapping relation is obtained; and the SAR image to be processed is input into a network to obtain a super-resolution result. Compared with the prior art, the method has the advantages that feature propagation can be effectively enhanced, the training speed can be increased, and a high-quality SAR imagesuper-resolution result can be obtained.

The invention provides a video space-time super-resolution implementation method and device, and the method comprises the steps of carrying out the edge enhancement of a video frame of a video, and obtaining an edge-enhanced video frame; inputting the plurality of edge-enhanced adjacent video frames into an optical flow estimation module in pairs to obtain a bidirectional optical flow; calculating the bidirectional optical flow to obtain an estimated optical flow, and inputting the estimated optical flow and the bidirectional optical flow into a bidirectional prediction module together to obtain a predicted optical flow; calculating the prediction optical flow and the corresponding video frame to obtain an intermediate frame for time domain super-resolution, and inserting the intermediate frame into a corresponding position in the video; and performing spatial domain super-resolution processing on the intermediate frame and the corresponding video frame through a cyclic super-resolution network to obtain a plurality of reconstructed frame; and circularly executing the steps until the space-time super-resolution of the whole video is completed. The invention has the beneficial effects that space-time joint super-resolution can be effectively carried out on the video, and the visual quality of the video is improved.

The invention discloses a residual instance regression super-resolution reconstruction method based on multistage dictionary learning, and the method comprises the following steps: generating a training set through high-resolution images, and establishing block pairs of low-resolution and high-resolution image blocks; extracting feature vectors of low-resolution image blocks, and learning a dictionary with strong representation ability by using K-SVD as an anchor point; performing the least square regression of low-resolution and high-resolution blocks in the block pairs through the dictionaryobtained via learning, and obtaining a linear mapping relation; estimating the high-resolution features, calculating a reconstruction error, and carrying out the mapping of the estimated high-resolution features and the reconstruction error while the further dictionary learning of the estimated high-resolution features; obtaining a group of residual regression devices after the L layer; carryingout the reconstruction through an inputted image and the obtained residual regression devices, and enabling the obtained high resolution features to be used for the reconstruction of a next layer; adding all estimated high-resolution image blocks and forming a high-resolution image through synthesis. The method is stronger in super-resolution capability, and can be used for the amplification of alow-resolution natural image.

The invention discloses an image denoising method and device, electronic equipment and an image super-division denoising method. The image denoising method comprises the following steps: respectivelyacquiring a first noiseless image and a second noiseless image with different image contents; performing alternate iterative training on the initial noise generation network and the discrimination network by using the first image and the second image to obtain a newly-built noise generation network; inputting the first image into a newly-built noise generation network, and outputting a third imagewith analog noise; training the initial image denoising network by using the first image and the third image as paired training samples to obtain a trained newly-built image denoising network, wherein the newly-built image denoising network is configured to convert the original image with noise into a noiseless newly-built image. According to the image denoising method, the image denoising network used for converting the low-resolution image into the high-resolution image can be obtained by training the asymmetric training set, and the image super-resolution effect is improved.

The invention discloses a super-resolution image texture optimization method and device. The method is realized on the basis of time-frequency feature extraction and mapping kernel learning, and comprises the following steps: 1) training a local texture block sample screener on the basis of local texture blocks with time domain and fractional frequency domain features, and screening samples with obvious texture features from external samples to serve as training samples of a mapping kernel training module; (2) learning a texture block optimization mapping kernel based on a local texture block of external sample learning, and (3) carrying out texture optimization on a to-be-optimized picture. According to the method and the device provided by the invention, the texture of the super-resolution image can be clearer and more natural.

The invention provides a video super-resolution method based on time attention and a cyclic feedback network. The characteristic that the contribution degrees of visual information provided by adjacent frames with different distances from a target frame to a super-resolution reconstruction effect are different, a feedback mechanism of a human visual system and a cyclic feedback guidance characteristic in a human new knowledge learning process are applied to a video super-resolution technology; a time attention module is adopted to learn an attention map of a video sequence on a time axis, so that the contribution of adjacent frames with different time degrees to a final reconstruction effect can be effectively distinguished; a video sequence is rearranged and then is subjected to cyclic feedback super-division by a cyclic feedback module, and finally a super-resolution network model is obtained, and the model has the characteristic of emphatically learning information with high contribution to super-division reconstruction and strong high-level feature learning ability, so that the video super-resolution effect is improved.

The invention discloses a super-resolution image reconstruction method based on multi-parallax attention module combination. The method comprises the following steps: 1) constructing a training sample set; 2) constructing a multi-parallax attention module network; 3) training the multi-parallax attention module network; 4) obtaining a trained multi-parallax attention module network model; and 5) obtaining a super-resolution reconstruction image result. According to the invention, the three-dimensional image super-resolution network model based on the multi-parallax module combination structure and the image smoothing loss function is constructed, and the existing image super-resolution network model is improved in a more reasonable and flexible manner, so that the super-resolution imaging quality is effectively improved, compared with the existing super-resolution image reconstruction technology, the super-resolution image reconstruction method has better anti-interference capability and higher super-resolution performance, and can provide richer detail information for further processing of the super-resolution reconstructed image.

The invention discloses a video super-resolution method based on a two-way cyclic convolution network, comprising: establishing a two-way cyclic network, including a forward cyclic sub-network and a backward cyclic sub-network in chronological order, and each cyclic sub-network starts from the bottom Contains an input sequence layer, two hidden sequence layers and an output sequence layer, each sequence layer includes multiple states, corresponding to video frames at different times; use three convolution operations to connect these states, including feedforward Convolution, circular convolution and conditional convolution to obtain a bidirectional circular convolutional network; send the training video to the established bidirectional circular convolutional network, and use the stochastic gradient descent algorithm to minimize the predicted and true high-resolution The mean square error between the videos, thereby iteratively optimizing the weight of the network, and obtaining the final two-way circular convolution network; inputting the low-resolution video sequence to be processed to the final two-way circular convolution network model, and obtaining the corresponding super-resolution results.

The invention discloses a light field image multi-magnification spatial super-resolution method and device. The method comprises the following steps: S1, training a single-image super-resolution network based on an EDSR structure; s2, using the trained single-image super-resolution network model to carry out single-image super-resolution with the magnification of m on all view angle images of the low-spatial-resolution light field image; s3, extracting a polar plane image before restoration from a single-image super-resolution result to form a polar plane image cube; s4, constructing a U-shaped polar plane image cube multi-magnification repair network based on an attention residual structure; s5, training the multi-magnification repair network of the polar plane image cube, repairing the information of the polar plane image cube by using the trained network model, and obtaining the polar plane image cube in which the geometric continuity of the space target in the visual angle dimension is recovered; and S6, reconstructing the repaired polar plane image cube into a light field image space super-resolution result with the multiplying power of m. According to the method provided by the invention, the super-resolution effect of the network is improved.

The invention discloses a face image super-resolution method based on an improved deep iterative collaborative network, and the method comprises the steps: 1), carrying out the early-stage data processing, and obtaining low-resolution face image data; 2) inputting a low-resolution face image into an image super-resolution sub-network, and processing the face image by a shallow feature extraction module to obtain shallow features; 3) the output of the shallow feature and prior information extraction sub-network in the last iteration process is sent to the image super-resolution sub-network to obtain high-resolution features, and a reconstruction module reconstructs the high-resolution features to obtain a high-resolution image; 4) inputting the high-resolution image into the prior information extraction sub-network, and outputting a face key point thermodynamic diagram, a semantic analysis diagram and intermediate layer features at the same time; and (5) repeating the steps (3) and (4), and carrying out iteration for N times to obtain final high-resolution image output. According to the method, the structure information of the face image is fully considered, the super-resolution effect of the face image is better, the parameter scale is smaller, and the time overhead is smaller.

The invention provides a distributed array SAR sparse representation three-dimensional imaging algorithm considering scene amplitude real value constraints. According to the algorithm, a sparse representation theory is adopted, a complex scene is represented as a combination of amplitude and phase, a scene amplitude real value constraint term is added into a regularized reconstruction model by using two pieces of priori information that the three-dimensional scene sparsity and the scene amplitude as real numbers, and the scene amplitude and phase are respectively estimated by using a quasi-Newton algorithm, thereby completing three-dimensional reconstruction more conforming to the actual situation. According to the method, amplitude real value constraints are added to the regularized reconstruction model, so that the measurement model is more reasonable, the actual situation is met, and high-resolution three-dimensional imaging with higher super-resolution capability and higher robustness can be realized.

The invention discloses an unsupervised image super-resolution fuzzy kernel estimation method and a terminal, and the method comprises the steps: obtaining an original input image, carrying out the augmentation of the original input image, and outputting a plurality of images; obtaining any image in the plurality of images, and performing down-sampling through an encoder to obtain a down-sampled image; according to the down-sampled image and the original input image, closing the block distribution between the down-sampled image and the original input image through a discriminator; performing up-sampling on the down-sampled image through a decoder to obtain the size of an original input image to obtain a reconstructed image; extracting blurred kernels of the plurality of images after the augmentation operation through an encoder, and obtaining a final blurred kernel after average processing. According to the method, the blurred kernel is estimated by learning the internal information ofthe image through the encoder, and the estimated blurred kernel is corrected through the feedback of the decoder, so that the accuracy of blurred kernel estimation is improved, and the super-resolution performance of the image in an unsupervised scene is improved.

The invention relates to a video processing method and device. The video processing method comprises the following steps: acquiring all video frames of a video; for each of the video frames, the following processing is performed: calculating a change image between a current video frame and an adjacent video frame, wherein the adjacent video frame is a video frame adjacent to the current video frame, calculating spatial attention information between the current video frame and the adjacent video frame based on the change image, obtaining the features for super-resolution of the current video frame based on spatial attention information between the current video frame and the adjacent video frame, and generating a high-resolution video frame for the current video frame based on the feature for super-resolution. According to the video processing method and device, the attention to the complementary region can be improved based on the spatial attention information, and the attention to the redundant region can be reduced so that the super-resolution performance is improved.

The present invention provides a low-resolution face super-resolution and recognition method based on facial prior knowledge, including S1: building a data set; S2: performing face super-resolution to obtain a mapping matrix; S3: constructing Feature extractor, respectively map and to the public space; S4: obtain the corresponding mapping matrix; S5: obtain the prior knowledge of the image, and obtain multiple super-resolution results; S6: respectively map and to the public space, and Assign its category to; S7: finally generate the result of face super-resolution and recognition. Through the technical solution of the present invention, the content of the present invention mainly includes two parts, one is to train the face pair data set composed of low resolution and high resolution; Space to train a nonlinear transformer with the goal of improving the accuracy of low-quality face image recognition.

The invention discloses a residual instance regression super-resolution reconstruction method based on multi-level dictionary learning. For the feature vector of the image block, use K-SVD to learn a dictionary with strong representation ability as the anchor point; use the learned dictionary to perform least squares regression on the low-resolution and high-resolution blocks in the block to obtain a linear mapping relationship; Estimate the high-resolution features, calculate the reconstruction error, and map the estimated high-resolution features with the reconstruction error while doing further dictionary learning; after the L layer, a set of residual regressions is obtained; use the input image and the obtained The residual regressor performs reconstruction, and the obtained high-resolution features are used for the reconstruction of the next layer; all estimated high-resolution image blocks are summed and calculated to synthesize a high-resolution image. The invention has stronger super-resolution capability and can be used for enlarging low-resolution natural images.

The present invention provides a Transformer-based end-to-end face super-resolution method, including S1: data preprocessing, obtaining image block sequences; S2: using convolutional neural network as an encoder to extract image local features; S3: Transformer-based The encoder module uses the self-attention mechanism to extract global remote features based on the sequence of image blocks; S4: Combines global and local features to implement an end-to-end face super-resolution method. Through the technical solution of the present invention, the content of the present invention mainly includes two parts, one is to process two-dimensional images, using the self-attention mechanism to extract the non-local long-range dependent information of the image sequence; the other is to use the local features extracted by the convolution operation at the same time , the two are combined as the input of the super-resolution decoder, the purpose is to reduce the complexity of model training by using the end-to-end learning method while enhancing the image features.

The present invention relates to a fine-grained scale image super-resolution method based on a non-local enhancement network. The method includes the following steps: Step A: Preprocessing the original high-resolution training image to obtain low-quality high-resolution images of different scales A dataset of image patch pairs consisting of image patches and original high-resolution training image patches; Step B: Use the image patch dataset to train a non-locally enhanced deep network; Step C: Input high-resolution images of low-quality test images into A deep network is used for reconstruction to obtain super-resolution results. This method uses a non-locally enhanced deep residual structure. By combining non-local operations with ordinary convolutions, it can effectively capture and utilize local and non-local image attributes for image super-resolution. Compared with existing super-resolution models, This method can significantly improve the performance of image super-resolution on fine-grained scales.