The invention relates to a textural feature-based no-reference image quality evaluation method. The method comprises the following specific steps of: dividing a graphic database into a training image set and a test image set; constructing images uder different scales for each image; solving normalized brightness images for the images under different scales; solving normalized gray-level co-occurrence matrixes in four directions for the normalized brightness images; calculating energy, entropies, contrast ratios and dependencies of the normalized gray-level co-occurrence matrixes to combine a feature vector; sending the feature vector and a subjective MOS score to a support vector machine to carry out training; and predicting test images by using the trained support vector machine so as to obtain an objective image quality evaluation result. According to the method, image quality evaluation is carried out by utilizing textural features; compared with the existing algorithms, the method is lower in calculation complexity and convenient to realize in real time, and has the effect of improving the prediction precision through extracting the textural features by utilizing the gray-level co-occurrence matrixes with a plurality of scales and a plurality of angles.

Textural feature-based no-reference image quality evaluation method

The invention discloses an image super-resolution reconstruction method based on deep learning, and aims to train low-resolution data by using deep learning technology in order to obtain a mapping function from low resolution to high resolution. The method comprises steps of: (1) downsampling a data set; (2) adding the convolutional activation results of different layers by using a residual principle; (3) classifying training data into a labeled class and a unlabeled class which correspond to two loss numbers respectively; and (4) integrating the two classes to obtain a final loss function. Any one low-resolution image is input into a trained neural network model such that the output of the neural network is a reconstructed super-resolution image. The method can effectively improve the quality of the obtained image without changing the hardware of an imaging system.

Image super-resolution reconstruction method based on deep learning

The invention provides a video insertion method, a storage medium and a terminal, so as to solve the problem that the video insertion effect is poor in the prior art. The method comprises the following steps: sequentially determining a current frame of a to-be-inserted frame video, a previous frame of the current frame and a next frame of the current frame; inputting the current frame of the to-be-inserted frame video, the previous frame of the current frame and the next frame of the current frame into a pre-generated video insertion frame model, wherein the video insertion frame model is generated by training a pre-set convolutional neural network model by a current frame of a training set, a previous frame of the current frame and a next frame of the current frame; frame insertion is performed on the to-be-inserted frame video by the video insertion frame model so as to obtain the video after frame insertion. According to the embodiment of the invention, better video insertion effectcan be achieved.

Video insertion method, storage medium and terminal

The invention discloses a super-resolution implementation method based on reconstruction optimization and a deep neural network. The method comprises the following specific steps of 1, establishing arelationship between a high-resolution image and a down-sampled image by means of a down-sampling method, and establishing a target function by means of a least square method; 2, utilizing a conjugategradient descent algorithm to conduct iterative optimization on the target function to obtain a high-resolution image based on a reconstruction optimization algorithm; 3, establishing and completinga three-layer full convolutional neural network through training; 4, inputting the high-resolution image obtained in the step 2 into the three-layer convolutional neural network to further improve theresolution of the image. By means of the method, good results can be achieved in both subjective evaluation and objective image quality evaluation.

Super-resolution implementation method based on reconstruction optimization and deep neural network

The invention discloses a similar image compression method based on residual compression sensing. The method comprises the steps of 1) collecting a group of similar images; 2) performing intelligent training on the similar images to extract a common template, subtracting a first image to be compressed from the common template to obtain a residual, training the residual through the K-SVD (K-singular value decomposition) algorithm to obtain a common dictionary and a measurement value coefficient, performing compression sensing measurement on the residual to obtain a measured residual value, extracting the nonzero terms of the measured value for quantization and entropy coding to obtain a compression code stream, and extracting the position information of the nonzero terms as the header information; 3) perform anti-entropy coding and inverse quantization on the compression code stream to obtain a nonzero term residual measured value, restoring a measurement matrix by combining the header information, according to the dictionary, reconstructing a residual image through the OMP (orthogonal matching pursuit) sparse coding algorithm, and adding the residual image to the common temperate to obtain a reconstructed image; 4) outputting the reconstructed image. The similar image compression method based on residual compression sensing can greatly reduce data storage, effectively improve the compression efficiency and achieve high reconstruction quality.

Similar image compression method based on residual compression sensing

The invention discloses a time-dimension video super-resolution method based on deep learning, which is mainly aimed at solving the problems of the prior art that reconstructed video image insert frames are poor in stability and low in precision. The technical key of the method is that fitting of a nonlinear mapping relationship between an original video image and a down-sampling video image is conducted through neural network training. The method comprises the following steps: 1) obtaining an original video image set and a down-sampling video image set and taking the original video image set and the down-sampling video image set as training samples of a neural network; 2) constructing a neural network model, and training parameters of the neural network through the training samples; and 3) taking any given video as a test sample, and inputting the test sample into the trained neural network model, wherein an output result of the neural network is a reconstructed video image. The calculating complexity of reconstruction of video image insert frames is reduced, and the stability and the precision of the reconstructed video image insert frames are improved. The method can be used for scene interpolation and animation making and also for time-domain insert frames of low-frame-rate videos.

Time-dimension video super-resolution method based on deep learning

The invention discloses an improved natural scene statistic model-based no-reference image quality evaluation method, and mainly aims at solving the problem that the evaluation carried out on noise images by computers does not accord with human eye perception. The method is realized through the following steps of: 1, giving a training set which comprises various types of noise images, and extracting characteristics of each noise image; 2, respectively training a classification model and a regression model by adopting a support vector machine according to the characteristics, noise types and subjective quality value of the noise images in the training set; 3, calculating characteristics of a test image; 4, calculating a probability that the test image belongs to each noise type according to the classification model and the characteristics of the test image; 5, calculating a quality value of the test image when the test image belongs to each noise type according to the regression model and the characteristics of the test image; 6, calculating a final quality of the test image; and 7, obtaining the quality evaluation of the test image according to the final quality value. According to the method, the evaluation result is in accordance with the human eye perception, so that the method can be used for the correct quality identification of natural images and remote sensing images.

Improved natural scene statistic model-based no-reference image quality evaluation method

The invention discloses an image compression method based on vision redundancy measurement. The image compression method mainly solves the problem that an existing image compression method is low in image compression efficiency. The image compression method includes the steps: (1) forecasting an image to be compressed through an autoregression model to obtain a forecasted image, (2) obtaining a residual image according to the image to be compressed and the forecasted image, and calculating local binary pattern values of pixels in the residual image, (3) counting the occurring probability of the same local binary pattern values in the neighbor areas around the pixels, and calculating the uncertainty of the image to be compressed, (4) calculating visual threshold values corresponding to the pixels according to the uncertainty to conduct redundancy removal operation on the image to be compressed to obtain the pre-processed image, and (5) compressing the pre-processed image to obtain a compressed image. According to the image compression method, vision redundancy information in the image can be effectively removed through the visual threshold values, the compression efficiency is promoted, and the image compression method can be used for transmitting online video, mobile phone picture-phone images and satellite remote sensing images.

Image compression method based on vision redundancy measurement

The invention discloses a sub-band-information-entropy-measure-based image quality evaluation method. With the method, a problem that evaluation of a noise image by a computer does not conform to the perception of the human eyes can be solved. The method comprises: step one, block discrete cosine transform is carried out on a to-be-tested image; step two, similarity frequency band coefficients of all sub block transform coefficients are combined; step three, the combined coefficients are reconstructed and the identical frequency band coefficients form reconstruction sets; step four, an information entropy of each reconstruction set coefficient is calculated; step five, the operations from the step to step four are carried out on a reference image corresponding to the to-be-tested image to obtain an information entropy of the reference image; step six, differences of information contents on all frequency bands of the to-be-tested image and the reference image are calculated; step seven, a weighted sum of all frequency band information content differences is calculated to obtain a quality value of the to-be-tested image; and step eight, according to the quality value, quality evaluation of the to-be-tested image is obtained. According to the invention, the obtained evaluation result conforms to the perception of the human eyes. The method can be used for identifying image transmission and image retrieving on the internet and measuring the quality of the image processing index.

Sub-band-information-entropy-measure-based image quality evaluation method

The invention discloses a characteristic dictionary-based no-reference image quality evaluation method, and mainly aims at solving the problem that the evaluation carried out on noise images by computers does not accord with human eye perception. The method is realized through the following steps of: 1, randomly selecting 60 natural images and calculating a local mode of each pixel in all the images; 2, clustering the local modes by adopting a K-means method, and forming a characteristic dictionary by a clustering center; 3, extracting characteristics of each training image in a training sample set on the basis of the characteristic dictionary; 4, training a regression model through support vector regression according to the characteristics of the training images and corresponding subjective quality values; 5, extracting characteristics of a test image on the basis of the characteristic dictionary; 6, inputting the characteristics of the test image into the regression model and predicting the image quality; and 7, obtaining the quality evaluation of the test image according to a quality value. According to the method, the evaluation result is in accordance with the human eye perception, so that the method can be used for identifying the quality of natural images, remote sensing images and the like, and measuring the performance of imaging equipment.

Wu Jinjian

Papers

Time-dimension video super-resolution method based on deep learning

Improved natural scene statistic model-based no-reference image quality evaluation method

Image compression method based on vision redundancy measurement

Sub-band-information-entropy-measure-based image quality evaluation method

Characteristic dictionary-based no-reference image quality evaluation method