31results about How to "Improve objective quality" patented technology

The invention provides a method for constructing a convolutional neural network for video coding fractional pixel interpolation, which comprises the following steps: images with different content andresolution are collected, and an original training data set containing data with different types and coding complexity is formed; preprocessing operation is performed on the original training data setto obtain training data conforming to the video coding inter-frame prediction fractional pixel interpolation characteristic; a deep convolutional neural network is built to obtain a convolutional neural network structure suitable for the video coding inter-frame prediction fractional pixel interpolation; the pre-processed data is input into a built-up convolutional neural network; meanwhile, theoriginal training data set is used as a corresponding true value to train the built-up convolutional neural network. According to the method, the convolutional neural network can be successfully trained; the fractional pixels obtained by using the trained convolutional neural network interpolation meet the requirement for video coding fractional pixel interpolation characteristic; and the method in the invention is used for performing fractional pixel interpolation so that the video coding efficiency can be improved.

The invention discloses an image interpolation method basing on a video object and the regional guidance. The particular process follows like that: an original image is divided and the position and the region of an interpolation point are determined; to the interpolation point in the internal of a region, a one-dimensional linear interpolation formula is adopted for the evaluation and a two-dimensional nonlinear interpolation formula is adopted for the evaluation of the interpolation point on other positions when the interpolation point is positioned between two horizontal pixels or two vertical pixels at the original image, to the interpolation points in other positions, the two-dimensional nonlinear interpolation formula is used for the evaluation is evaluated when the interpolation point is positioned neither between the two horizontal pixels of the original image nor between the two vertical pixels of the original image; the obtained values of each pixel point are endowed to the pixel on the position of an interpolation point waiting to be interpolated to finish the image interpolation. The image interpolation method basing on the video object and the regional guidance is applicable to the transforming of the resolution of a video object or a whole image.

The invention belongs to the technical field of optical sensor image processing, and discloses a multi-sensor image fusion method and a system based on gradient domain guiding filter. The method smoothes a source image by using a mean filter, removes a small structure in the source image, and decomposes the source image to obtain a base layer and a detail layer of the source image. Laplace high-pass filter and Gaussian low-pass filter are used to filter the source image, and the salient image is obtained. The weight map of the corresponding source image is obtained by comparing the salient pixel size of the source image. The source image is used as the guide image and the weight map is decomposed by GDGF to get the basic layer and the detail layer of the weight map respectively. Accordingto certain fusion rules, the pixels corresponding to the base layer and the detail layer are fused by using the optimized weight map base layer and the detail layer, and the fused base layer and the detail layer are merged to obtain the fused image. The invention can effectively save the detail information in the source image, greatly improves the subjective and objective quality of the fused image, and has robustness to image registration.

The invention belongs to the technical field of optical image processing and discloses an AGF-based multi-focus image fusion method and system. The method comprises the steps of firstly smoothing an input image by using joint bilateral filtering, and alternately using a source image and a filtered image as the input image and a guide image of bilateral filtering; performing filtering processing onthe bilaterally filtered image by using median filtering to obtain a basic layer and a detail layer of the source image; calculating gradient energy of neighborhood windows of pixels of the basic layer and the detail layer of the source image, establishing a decision matrix according to the gradient energy of the neighborhood windows of the pixels of the basic layer and the detail layer, and fusing the pixels corresponding to the basic layer and the detail layer according to a certain fusion rule. The accuracy of judging a focus region in the source image can be effectively improved and the quality of a fused image can be greatly improved.

Provided is an improved sample adaptive offset filtering method based on histogram analysis. The method comprises the steps of analyzing histogram distribution according to gray values of sample values in coding tree blocks of a reconstruction frame, classifying the coding tree blocks according to the histogram distribution, dividing adaptive sample offset filtering into a narrow coding tree block mode, a wide coding tree block mode, a double-center coding tree block mode and a default mode, respectively calculating the optimal rate distortion cost values under different classification modes, selecting the mode corresponding to the minimum rate distortion cost value as a truly-adopted band filtering mode and coding and transmitting a corresponding band starting position and an offset value. Three more accurate and more efficient filtering classification methods are newly added according to the characteristics of coding tree block histogram distribution so as to improve the accuracy of the sample adaptive offset filtering method, and the subjective and objective quality of videos can be effectively improved under the condition of same code rate.

The invention discloses a stereoscopic video whole frame loss error hiding method based on structural similarity. The stereoscopic video whole frame loss error hiding method based on the structural similarity is effectively combined with subjective perceptions from human eyes to picture structure information, respectively uses a motion compensation forecasting method or a parallax compensation forecasting method to recover errors aiming at different macro block reference modes of macro blocks in a dropped frame by judging the macro block reference modes of reference image frames of a moment before the dropped frame, and further is intensively combined with the subjective perceptions from the human eyes to picture structure similarity due to the fact that time domain pertinence and pertinences among viewpoints of traditional stereoscopic video are fully considered, and therefore not only objective quality of recovering of the dropped frame can be improved, but also subjective quality of the recovering of the dropped frame can be enabled to be close to the perceptions of the human eyes.

The invention belongs to the technical field of image encoding and decoding and discloses a differential image compression perception reconfiguration method based on multi-hypothesis weighting and an intelligent terminal. A cross-block-based block compression sensing process is used to sample an original image to obtain a measurement value. A non-local means fully differential iterative reconstruction algorithm is used to carry out iterative reconfiguration processing on the obtained measurement value, and an initial image reconstruction value is obtained. The multiple-hypothesis set acquisition processing is carried out on a current reconstruction value, and the optimized filtering processing is carried out on an obtained multi-hypothesis set to eliminate an inferior hypothesis. The optimized multi-hypothesis set is processed by using a weight estimation model based on an elastic network, a multi-hypothesis weight matrix is obtained, the weighted summation processing is carried out on the multiple hypotheses to obtain side information, and a more accurate iteration initial value is provided for subsequent iterations. According to the method and the intelligent terminal, the spatial correlation of the image is effectively used, the multi-hypothesis weighting processing is used to effectively relieve an over-smoothing problem of a past reconstruction algorithm, and the image reconstruction quality is greatly improved.

The invention discloses an H.265-baased rate control method, system and device. The method comprises the following steps: judging whether a current frame is a scene switching frame; if yes, setting the current frame as an I frame, updating a quantization parameter according to the relation between the quantization parameter and a Lagrangian multiplier method coefficient, and carrying out target bit distribution on new GOP and new frame layers; if no, continuing to carry out target bit distribution on the new GOP and the new frame layers; calculating the quantization parameter by virtue of an R-Q model and a frame layer target bit distribution result; carrying out rate-distortion optimization according to the calculated quantization parameter, and selecting an optimal coding mode; coding inthe optimal coding mode; if the current frame is the last frame, ending the operation; and otherwise, updating model parameters. The system comprises modules (1-11). The device comprises a memory anda processor for executing the method. According to the method, the system and the device, the objective quality of scene switching videos can be improved; and the method, the system and the device can be widely applied to the video coding field.

The invention provides a method for dividing and rebuilding a video coding predictive residue block, and belongs to the technical field of video coding in signal treatment. The invention aims to reduce the blocking effect due to the prior DCT transformation matrix and solve the problem that the value of the matrix cannot be adjusted. The division method comprises a predictive residue block classification step, a transverse division step of first, third and second multi-channel filter groups, a classification step of transverse division coefficient matrixes and a longitudinal division step of the first, the third and the second multi-channel filter groups; and the rebuilding method comprises a classification step of the division coefficient matrixes of the predictive residue block, a longitudinal rebuilding step of the first, the third and the second multi-channel filter groups; a classification step of the transverse division coefficient matrixes and a transverse rebuilding step of the first, the third and the second multi-channel filter groups. The method has the advantages of effectively removing correlation, reducing the blocking effect due to the size mismatching of the DCT transformation matrix and the predictive residue block, and improving subjective and objective qualities of coding.

The invention discloses a stereo video bitrate control method based on a binocular vision characteristic. The method is characterized in that bitrate control is conducted on a stereo image group layer, a stereo image pair layer and a frame layer respectively; in the stereo image group layer, the bit number of each stereo image group and a quantization parameter of a key frame are calculated; in the stereo image pair layer, the target bit number of each stereo image pair is calculated according to the surplus bit number and the buffer area saturation; in the frame layer, a stereo index bitrate and quantization parameter model is built according to the human eye binocular vision masking effect to optimize a stereo rate distortion model to enable more target bits to be distributed in a left image and fewer target bits to be distributed in a right image, and finally the quantization parameters of non key frames can be determined. The method has the advantage of effectively improving stereo video objective quality and rate distortion performance.

The invention discloses a three-dimensional image reduction method which comprises the following steps: calculating gradients in the directions of 0 degree, 45 degrees and 135 degrees according to formula (1), formula (2) and formula (3), wherein weighting coefficients are exhaustively chosen from 10 groups of coefficients according to PSNR (peak signal to noise ratio); and calculating interpolations in the directions of 0 degree, 45 degrees and 135 degrees by adopting the principle of interpolating in the direction with the smallest gradient, wherein formula (1), formula (2) and formula (3) refer to the patent specification. As the three-dimensional image reduction method adopts the principle of interpolating in the direction with the smallest gradient, image objective quality and particularly the subjective quality are improved.

The invention belongs to the field of ocean engineering, and particularly relates to an underwater image enhancement method based on unsupervised color correction. According to the method, the limitation that the channel B is used as a protruding channel is improved, so that the color distortion problem of processing the non-blue-color-cast underwater image is effectively solved; The method comprises an RGB color model, a contrast correction method, an HIS color model and an underwater image enhancement method based on unsupervised color correction. Compared with a common underwater image enhancement method, the underwater image enhancement method has the advantages that the underwater image enhancement method is improved into a self-adaptive method and has good processing effects on different color cast; The image is effectively balanced, color cast is eliminated, the illumination is improved, and the true color is increased. The method is easy to implement, small in calculated amountand high in reliability, the feasibility of underwater optical image recognition is improved, and the method has positive significance for development of unmanned underwater vehicles in the aspects of underwater operation tasks and the like in the future.

The invention relates to a ternary-representation-based image predictive coding method, which is particularly suitable for the compression processing of static images. The method comprises the following steps of: performing wavelet transform and wavelet coefficient quantization on an image, representing each wavelet coefficient by using a ternary number and scanning a ternary wavelet coefficient plane; selecting a nearest-neighbor coefficient of a symbol currently to be coded as a prediction coefficient, defining an importance state function, an importance state direction weighting function and an importance state and function expression of the prediction coefficient for the characteristics of three symbols, and establishing a high-efficiency prediction model; and calculating a predicted value according to the prediction model, transmitting the symbol currently to be coded into a corresponding arithmetic coder for entropy coding, and resetting an initial value of the arithmetic coder between frequency bands. By the method, the high-efficiency predictive coding of wavelet coefficients is realized, and the objective quality of image recovery is effectively improved.