WO2018119565A1 - 一种桶形畸变图像的矫正重构方法及装置 - Google Patents
一种桶形畸变图像的矫正重构方法及装置 Download PDFInfo
- Publication number
- WO2018119565A1 WO2018119565A1 PCT/CN2016/112068 CN2016112068W WO2018119565A1 WO 2018119565 A1 WO2018119565 A1 WO 2018119565A1 CN 2016112068 W CN2016112068 W CN 2016112068W WO 2018119565 A1 WO2018119565 A1 WO 2018119565A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- image
- resolution
- low
- training
- low resolution
- Prior art date
Links
- 238000000034 method Methods 0.000 title claims abstract description 43
- 239000011159 matrix material Substances 0.000 claims abstract description 61
- 238000011084 recovery Methods 0.000 claims abstract description 29
- 238000012549 training Methods 0.000 claims description 162
- 238000013507 mapping Methods 0.000 claims description 20
- 238000012545 processing Methods 0.000 claims description 12
- 238000004364 calculation method Methods 0.000 claims description 5
- 238000006243 chemical reaction Methods 0.000 claims description 3
- 239000000284 extract Substances 0.000 claims description 3
- 238000000605 extraction Methods 0.000 claims description 2
- 238000003702 image correction Methods 0.000 abstract description 4
- 238000012937 correction Methods 0.000 description 9
- 238000010586 diagram Methods 0.000 description 6
- 239000012634 fragment Substances 0.000 description 4
- 238000012805 post-processing Methods 0.000 description 3
- 230000000694 effects Effects 0.000 description 2
- 241000226585 Antennaria plantaginifolia Species 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000000704 physical effect Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
Images
Classifications
-
- G06T5/80—
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20016—Hierarchical, coarse-to-fine, multiscale or multiresolution image processing; Pyramid transform
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
Definitions
- the invention belongs to the technical field of computers, and in particular relates to a method and a device for correcting and reconstructing a barrel distortion image.
- the wide-angle lens has a wide viewing angle range and a long depth of field. It is often used to shoot large scenes or towering buildings. At the same time, as the wide-angle lens is used more and more, the nonlinear lens distortion of the camera is also affected. More people's attention.
- Lens distortion can be generally divided into three types: pincushion distortion, barrel distortion and linear distortion.
- the barrel distortion is the distortion phenomenon of the barrel image expansion caused by the physical properties of the lens in the lens and the lens group structure.
- the research directions related to barrel distortion mainly include how to improve the quality of wide-angle lens images and how to improve the accuracy of correcting distortion process, and rarely study how to improve the resolution of distorted images.
- Some researchers have proposed to use a reconstruction-based
- the super-resolution method is used to improve the resolution of the fisheye camera. However, this method requires multiple frames of images to be acquired, which is inefficient, and in most super-resolution applications, the reconstruction-based method is generally less effective than the learning-based method.
- the object of the present invention is to provide a method and a device for correcting and reconstructing a barrel distortion image, which aims to solve the problem of correcting and reconstructing an effective barrel distortion image due to the prior art, resulting in correction of the barrel distortion image.
- the structure is inefficient and corrects the problem of poor image quality after reconstruction.
- the present invention provides a method for correcting and reconstructing a barrel distortion image, the method comprising the steps of:
- the present invention provides a corrective reconstruction device for a barrel distortion image, the device comprising:
- a first distortion recovery module configured to receive a low resolution barrel distortion image input by a user, and perform a distortion recovery operation on the low resolution barrel distortion image to generate a corresponding low resolution image
- a training data acquisition module configured to extract image features of the low-resolution image, and search for a dictionary atom corresponding to each image feature of the low-resolution image in a pre-trained low-resolution sparse dictionary, and pre-train Obtaining a projection matrix corresponding to the dictionary atom in a plurality of projection matrices;
- a first reconstruction module configured to reconstruct, according to the image feature of the low resolution image and the projection matrix, a first high resolution image feature corresponding to the low resolution barrel distortion image
- a second reconstruction module configured to reconstruct, according to the first high resolution image feature and the pre-trained plurality of coefficient matrices, a second high resolution image feature corresponding to the low resolution barrel distortion image
- an image output module configured to generate, according to the second high resolution image feature, a high resolution image corresponding to the low resolution barrel distortion image, and output the high resolution image.
- the invention firstly performs distortion recovery on a low resolution barrel distortion image input by a user, generates a low resolution image, extracts image features of the low resolution image, and pre-trains low resolution sparse words. Searching for dictionary atoms corresponding to each image feature of the low-resolution image, searching for a projection matrix corresponding to the dictionary atom in a plurality of pre-trained projection matrices, and then, according to image features and projection matrices of the low-resolution image Performing the first reconstruction of the low-resolution barrel distortion image to generate the first high-resolution image feature, and performing the second reconstruction according to the pre-trained plurality of coefficient matrices to generate the second high-resolution image feature Finally, the low-resolution barrel distortion image is corrected and reconstructed, so that the image details corrected by the barrel distortion image are restored by the first reconstruction, and the image is further improved by the second reconstruction.
- the resolution effectively improves the efficiency and image quality of the barrel distortion image correction reconstruction.
- FIG. 1 is a flowchart of an implementation of a method for correcting and reconstructing a barrel distortion image according to Embodiment 1 of the present invention
- FIG. 2 is a flowchart showing an implementation process of a training process in a method for correcting and reconstructing a barrel distortion image according to Embodiment 2 of the present invention
- FIG. 3 is a diagram showing an example of a mapping relationship in a nonlinear distortion model in a method for correcting and reconstructing a barrel distortion image according to Embodiment 2 of the present invention
- FIG. 4 is a diagram showing an example of a mapping relationship corresponding to a high-resolution barrel distortion training image in a nonlinear distortion model in a method for correcting and reconstructing a barrel distortion image according to a second embodiment of the present invention
- FIG. 5 is a schematic structural diagram of a device for correcting and reconstructing a barrel distortion image according to Embodiment 3 of the present invention.
- FIG. 6 is a schematic structural diagram of a module for performing a training operation in a correction and reconstruction apparatus for a barrel distortion image according to Embodiment 4 of the present invention.
- Embodiment 1 is a diagrammatic representation of Embodiment 1:
- FIG. 1 is a flowchart showing an implementation of a method for correcting and reconstructing a barrel distortion image according to Embodiment 1 of the present invention. For convenience of description, only parts related to the embodiment of the present invention are shown, which are described in detail as follows:
- step S101 a low resolution barrel distortion image input by the user is received, and a distortion recovery operation is performed on the low resolution barrel distortion image to generate a corresponding low resolution image.
- the barrel distortion image when a barrel distortion image with a lower resolution input by the user is received, the barrel distortion image may be subjected to distortion recovery (or correction) to generate a corresponding low resolution image.
- the polar coordinate system of the low-resolution barrel distortion image may be established first, and the polar coordinate radius in the polar coordinate system is obtained according to the training.
- the mapping relationship is converted into a new polar coordinate radius, and the new polar coordinate radius is set as the polar coordinate radius of the image after distortion recovery, thereby generating an image after distortion recovery, that is, a low resolution image.
- step S102 the image features of the low-resolution image are extracted, and the dictionary atoms corresponding to each image feature of the low-resolution image are searched in the pre-trained low-resolution sparse dictionary, and the plurality of projection matrices are pre-trained. Get the projection matrix corresponding to the dictionary atom.
- the low resolution sparse dictionary and the plurality of projection matrices are all trained data during the training, and each dictionary atom corresponds to a projection matrix in the low resolution sparse dictionary.
- the dictionary element corresponding to the image feature of the low-resolution image is a dictionary atom in the low-resolution dictionary that is closest to the feature of the image.
- the low-resolution image when extracting image features of a low-resolution image, may be first filtered, and each low-resolution image may be filtered to obtain a plurality of filtered images, and then, a grid is adopted. The image fragments of the filtered images are extracted, and the image fragments of the filtered image at the same position are combined to obtain image features of the low-resolution image.
- step S103 the first high resolution image feature corresponding to the low resolution barrel distortion image is reconstructed according to the image feature of the low resolution image and the projection matrix.
- the resolution image feature is the image feature of the high resolution image obtained after the first reconstruction.
- the first reconstruction can be performed using the following formula:
- y T,j P k x T,j , where y T,j is the jth first high resolution image feature, x T,j is the image feature of the jth low resolution image, P k is The projection matrix corresponding to k dictionary atoms, the kth dictionary atom is the dictionary atom closest to the image feature distance of the jth low resolution image in the low resolution sparse dictionary.
- step S104 the second high resolution image feature corresponding to the low resolution barrel distortion image is reconstructed according to the first high resolution image feature and the plurality of coefficient matrices pre-trained.
- this reconstruction is referred to as a second reconstruction, and the information loss degree of the edge portion and the middle portion of the barrel distortion image in the distortion recovery process is different, but in the step
- the first reconstruction performed in S103 does not take into account the image characteristics of the barrel distortion image, so a post-processing is also required, that is, the second reconstruction of this step is performed.
- the second high resolution image feature is an image feature of the second high resolution image obtained by the second reconstruction.
- the first high-resolution image features may be classified, and for each type of image features after classification, the coefficient matrix corresponding to the image features is used for the second reconstruction, specifically Ground, the calculation formula for the second reconstruction can be:
- y R,c H c y T,c , where y R,c represents the second high-resolution image feature reconstructed when the first high-resolution feature is the c-type image feature, y T,c represents c for the first image feature a first class of high resolution features, H c is the coefficient matrix of the image features corresponding to class c.
- the first high-resolution image features are classified according to the distance from the first high-resolution image feature to the center of the first high-resolution image, so that the features of different regions of the image are selected according to the graphical characteristics of the barrel-shaped distortion pattern.
- the second reconstruction is performed by different coefficient matrices, which effectively improves the image resolution and image quality after the second reconstruction.
- step S105 a high resolution image corresponding to the low resolution barrel distortion image is generated according to the second high resolution image feature, and the high resolution image is output.
- all of the second high resolution image features are combined to produce a corrected reconstructed high resolution image.
- the low-resolution barrel distortion image input by the user is subjected to distortion recovery, first reconstruction, and second reconstruction, and finally a high-resolution image is generated, which is passed in the first reconstruction.
- the pre-trained low-resolution sparse dictionary and projection matrix restore the details of the image after distortion correction.
- the pre-trained coefficient matrix is used to perform post-processing on different regions of the restored image. Further improve the resolution and quality of the image, and solve the problem that the edge loss and the middle part of the barrel distortion image have different information loss degree during the distortion recovery process, thereby effectively improving the efficiency and correction of the barrel distortion image correction reconstruction.
- the resolution and quality of the reconstructed image is used to perform post-processing on different regions of the restored image.
- Embodiment 2 is a diagrammatic representation of Embodiment 1:
- FIG. 2 is a flowchart showing an implementation process of a training process in a method for correcting and reconstructing a barrel distortion image according to Embodiment 2 of the present invention, which is described in detail as follows:
- step S201 a plurality of high resolution training images stored in advance are processed to obtain a low resolution barrel distortion training image for training a low resolution sparse dictionary, a projection matrix, and a coefficient matrix.
- the existing plurality of high-resolution images are used as training images in the training process.
- these training images are referred to herein as high-resolution training images, and similarly, low.
- the resolution barrel distortion training image is a low resolution resolution barrel distortion image obtained by processing the high resolution training image for training a low resolution sparse dictionary, a projection matrix, and a coefficient matrix.
- the high resolution training image is processed to obtain a low resolution barrel distortion processed image by the following steps:
- the high-resolution training image is distorted to generate a high-resolution barrel distortion training image.
- a polar coordinate system may be first established for a high-resolution training image, a polar coordinate radius in the polar coordinate system is obtained, and a pole of a high-resolution barrel distortion training image is obtained according to a formula of a nonlinear distortion model.
- the coordinate radius which in turn generates a high-resolution barrel distortion training image.
- the formula of the nonlinear distortion model is Where r is the polar coordinate radius of the high resolution training image, r new is the polar coordinate radius of the high resolution distortion training image, and a i and n are preset parameters.
- n 2
- r new a 1 r + a 2 r 2
- the operation at this time is relatively simple
- n 2
- a 1 is At 1 o'clock
- FIG. 3 represents a non-distorted image (here, it can be considered as a high-resolution training image)
- the polar coordinate radius value the ordinate represents the polar radius value of the distorted image (here, can be regarded as a high-resolution barrel distortion image)
- r max is the maximum value of the non-distorted image polar coordinate radius.
- the high-resolution barrel distortion training image when the high-resolution barrel distortion training image is processed, the high-resolution barrel distortion training image may be downsampled, and the image obtained after down-sampling is subjected to interpolation processing.
- the interpolation process is performed by using an existing bicubic interpolation (bicubic interpolation) method in the interpolation process, thereby improving the image quality of the low resolution distortion training image to some extent.
- bicubic interpolation bicubic interpolation
- step S202 a distortion recovery operation is performed on the low resolution barrel distortion training image to generate a corresponding low resolution training image.
- the low resolution barrel distortion training image is restored (or corrected) to a low resolution training image.
- performing a distortion recovery operation on the low-resolution barrel distortion training image, and generating a corresponding low-resolution training image can be implemented by the following steps:
- the mapping relationship between the polar coordinate radius of the low-resolution barrel distortion training image and the polar coordinate radius of the low-resolution training image is obtained.
- a mapping relationship between the polar coordinate radius of the distorted image and the non-distortion image can be obtained in the nonlinear distortion model, and the non-distorted image can be deformed into a distorted image by using the mapping relationship, or The distortion image is corrected to a non-distorted image, so the mapping relationship between the polar coordinate radius of the low-resolution barrel distortion training image and the polar coordinate radius of the low-resolution training image can be obtained according to the nonlinear distortion model.
- the mapping relationship of the polar coordinate radius is
- r 1 is the polar coordinate radius of the low resolution barrel distortion training image
- r 2 is the polar coordinate radius of the low resolution training image
- the low resolution training image may be generated according to the polar coordinate radius.
- step S203 image features of the low resolution training image are extracted, and image features of the high resolution training image are extracted.
- all low-resolution training images may be filtered by using a preset number of high-pass filters, and each low-resolution training image may obtain a corresponding number of filtered images, which may be obtained by meshing methods. Extracting multiple image fragments and combining the image fragments at the same position in these images can obtain image features for each position of these images. Then, the mesh feature can be directly used to extract the image features at the same position of the high-resolution training image to obtain the image features of the high-resolution training image, and the image features of the low-resolution training image and the high-resolution training image can be seen. There is a correspondence between the image features.
- step S204 the low-resolution sparse dictionary and the projection matrix are calculated according to the image features of the low-resolution training image and the image features of the high-resolution training image, and the image features and the high-resolution training image of the image are trained according to the low resolution.
- the image features and the low-resolution sparse dictionary reconstruct the high-resolution reconstructed image features corresponding to the low-resolution barrel distortion training image.
- the low-resolution sparse dictionary and the projection matrix can be calculated by the following steps:
- the formula for calculating the low resolution sparse dictionary may be Where ⁇ is sparse representation coefficient matrix, x is image feature of low resolution training image, D L is low resolution sparse dictionary, ⁇ is preset weighting factor, and D L expression formula can be d L,k is the kth dictionary atom in the low resolution sparse dictionary, and K is the preset constant.
- the nearest neighbor feature of each dictionary atom in the image features of the low resolution training image is calculated by using the existing KNN algorithm (K nearest neighbor algorithm), the first image feature neighborhood is obtained, and the image is trained according to the low resolution.
- K nearest neighbor algorithm K nearest neighbor algorithm
- the correspondence between the image features and the image features of the high-resolution training image obtains the second image feature neighborhood, thereby effectively improving the computational efficiency of the nearest neighbor feature.
- the calculation formula of the projection matrix may be Where P k is the projection matrix corresponding to the kth dictionary atom, N H,k is the second image feature neighborhood, and N L,k is the first image feature neighborhood.
- the high-resolution reconstructed image feature is an image feature of the high-resolution image obtained after the training image is reconstructed.
- the reconstruction formula of the high-resolution reconstructed image feature may be: Where y H,i is the ith high resolution reconstructed image feature.
- step S205 a coefficient matrix is calculated based on the image features of the high resolution reconstructed image feature and the high resolution training image.
- the image features of the high-resolution reconstructed image and the high-resolution training image may be separately classified, and the coefficient matrix corresponding to each type of image feature after classification is calculated, specifically, the calculation formula of the coefficient matrix Can be
- H c is the coefficient matrix corresponding to the c-type image feature
- y H,c is the high-resolution reconstructed image feature of the c-type image feature
- y c is the image of the high-resolution training image of the c-type image feature feature.
- the high-resolution training image is advanced by a preset nonlinear distortion model.
- the line distortion is obtained, and the low-resolution barrel distortion training image is obtained, and the low-resolution barrel distortion image distortion is reconstructed and reconstructed, and the low-resolution sparse dictionary and the projection matrix corresponding to each dictionary element are trained.
- the coefficient matrix corresponding to each type of image feature effectively improves the effect of correcting and reconstructing the low-resolution barrel distortion image in the first embodiment.
- Embodiment 3 is a diagrammatic representation of Embodiment 3
- FIG. 5 is a diagram showing the structure of a correction and reconstruction apparatus for a barrel distortion image according to Embodiment 3 of the present invention. For the convenience of description, only parts related to the embodiment of the present invention are shown, including:
- the first distortion recovery module 51 is configured to receive a low-resolution barrel distortion image input by the user, and perform a distortion recovery operation on the low-resolution barrel distortion image to generate a corresponding low-resolution image;
- the training data obtaining module 52 is configured to extract image features of the low-resolution image, and search for the dictionary atom corresponding to each image feature of the low-resolution image in the pre-trained low-resolution sparse dictionary, which is pre-trained Obtain a projection matrix corresponding to the dictionary atom in the projection matrix;
- a first reconstruction module 53 configured to reconstruct a first high-resolution image feature corresponding to the low-resolution barrel distortion image according to the image feature and the projection matrix of the low-resolution image;
- a second reconstruction module 54 configured to reconstruct a second high-resolution image feature corresponding to the low-resolution barrel distortion image according to the first high-resolution image feature and the plurality of coefficient matrices that are pre-trained;
- the image output module 55 is configured to generate a high resolution image corresponding to the low resolution barrel distortion image according to the second high resolution image feature, and output the high resolution image.
- the low-resolution barrel distortion image input by the user is subjected to distortion recovery, first reconstruction, and second reconstruction, and finally a high-resolution image is generated, which is passed in the first reconstruction.
- the pre-trained low-resolution sparse dictionary and projection matrix restore the details of the image after distortion correction.
- the pre-trained coefficient matrix is used to perform post-processing on different regions of the restored image. , further improve the resolution and quality of the image, and solve the edge of the barrel distortion image
- the difference in the degree of information loss between the partial and intermediate parts during the distortion recovery process effectively improves the efficiency of the image correction reconstruction of the barrel distortion and corrects the resolution and quality of the reconstructed image.
- Embodiment 4 is a diagrammatic representation of Embodiment 4:
- FIG. 6 is a block diagram showing a module structure for performing a training operation in a correction and reconstruction apparatus for a barrel distortion image according to Embodiment 4 of the present invention. For convenience of description, only parts related to the embodiment of the present invention are shown, including:
- the training image processing module 61 is configured to process the plurality of high-resolution training images stored in advance to obtain a low-resolution barrel distortion training image for training the low-resolution sparse dictionary, the projection matrix, and the coefficient matrix;
- the second distortion recovery module 62 is configured to perform a distortion recovery operation on the low resolution barrel distortion training image to generate a corresponding low resolution training image;
- a feature extraction module 63 configured to extract image features of the low-resolution training image, and extract image features of the high-resolution training image
- the training reconstruction module 64 is configured to calculate a low-resolution sparse dictionary and a projection matrix according to the image features of the low-resolution training image and the image features of the high-resolution training image, and simultaneously train the image features and high resolution of the image according to the low resolution. Rate the image features of the training image and the low resolution sparse dictionary to reconstruct the high resolution reconstructed image features corresponding to the low resolution barrel distortion training image;
- the coefficient matrix calculation module 65 is configured to calculate a coefficient matrix according to the image features of the high resolution reconstructed image feature and the high resolution training image.
- the training image processing module 61 further includes an image distortion module 611 and an image resolution processing module 612, wherein:
- the image distortion module 611 is configured to perform distortion on the high-resolution training image according to the preset nonlinear distortion model to generate a high-resolution barrel distortion training image;
- the image resolution processing module 612 is configured to process the high resolution barrel distortion training image, Generate a low resolution barrel distortion training image.
- the second distortion recovery module 62 further includes a mapping relationship obtaining module 621 and an image conversion module 622, wherein:
- the mapping relationship obtaining module 621 is configured to obtain a mapping relationship between a polar coordinate radius of the low-resolution barrel distortion training image and a polar coordinate radius of the low-resolution training image according to the nonlinear distortion model;
- the image conversion module 622 is configured to convert the low resolution barrel distortion training image into a low resolution training image according to the mapping relationship.
- the high-resolution training image is distorted by a preset nonlinear distortion model, and a low-resolution barrel distortion training image is obtained, and the low-resolution barrel distortion image distortion is reconstructed.
- the first reconstruction and other operations, training to obtain a low-resolution sparse dictionary, a projection matrix corresponding to each dictionary element, and a coefficient matrix corresponding to each type of image feature, thereby effectively improving the low-resolution barrel distortion image in the third embodiment Perform the effect of correcting the reconstruction.
- each unit of the correction and reconstruction apparatus of the barrel distortion image may be implemented by a corresponding hardware or software unit, and each unit may be an independent software and hardware unit, or may be integrated into a soft and hardware unit. This is not intended to limit the invention.
- each unit reference may be made to the description of each step in the foregoing Embodiment 2, and details are not described herein again.
Abstract
本发明适用计算机技术领域，提供了一种桶形畸变图像的矫正重构方法及装置，该方法包括：对用户输入的低分辨率桶形畸变图像执行畸变恢复操作，生成对应的低分辨率图像，提取低分辨率图像的图像特征，在低分辨率稀疏字典中查找低分辨率图像的每个图像特征对应的字典原子，在多个投影矩阵中获取字典原子对应的投影矩阵，重构低分辨率桶形畸变图像对应的第一高分辨率图像特征，重构低分辨率桶形畸变图像对应的第二高分辨率图像特征，生成并输出低分辨率桶形畸变图像对应的高分辨率图像，从而不仅在第一次重构中恢复桶形畸变图像的细节，而且在第二次重构中进一步提高图像的分辨率，有效地提高了桶形畸变图像矫正重构的效率和图像质量。
Description
本发明属于计算机技术领域，尤其涉及一种桶形畸变图像的矫正重构方法及装置。
广角镜头具有较广的视角范围、较长的景深，常用来拍摄广阔的大场面或高耸入云的建筑物等，同时，随着广角镜头越来越多地被使用，相机的非线性镜头畸变也受到更多人的关注。
镜头畸变一般可以分为枕形畸变、桶形畸变和线性畸变三类，其中的桶形畸变是有镜头中透镜物理性能以及镜片组结构引起的成像画面呈桶形膨胀状的失真现象，目前，与桶形畸变相关的研究方向主要包括如何提升广角镜头拍摄图像的质量以及如何提高矫正畸变过程的准确率，而很少研究如何提高畸变图像的分辨率，有研究人员提出，使用一个基于重构的超分辨率方法来提高鱼眼相机的分辨率，然而该方法需要获取多帧的图像，效率较低，且在多数超分辨率应用中，基于重构的方法普遍比基于学习的方法效果差。
发明内容
本发明的目的在于提供一种桶形畸变图像的矫正重构方法及装置，旨在解决由于现有技术无法提供一种有效的桶形畸变图像的矫正重构方法，导致桶形畸变图像矫正重构的效率低且矫正重构后的图像质量较差的问题。
一方面，本发明提供了一种桶形畸变图像的矫正重构方法，所述方法包括下述步骤：
接收用户输入的低分辨率桶形畸变图像，并对所述低分辨率桶形畸变图像执行畸变恢复操作，生成对应的低分辨率图像；
提取所述低分辨率图像的图像特征，并在预先训练好的低分辨率稀疏字典中查找所述低分辨率图像的每个图像特征对应的字典原子，在预先训练好的多个投影矩阵中获取所述字典原子对应的投影矩阵；
根据所述低分辨率图像的图像特征和所述投影矩阵，重构所述低分辨率桶形畸变图像对应的第一高分辨率图像特征；
根据所述第一高分辨率图像特征和预先训练好的多个系数矩阵，重构所述低分辨率桶形畸变图像对应的第二高分辨率图像特征；
根据所述第二高分辨率图像特征，生成所述低分辨率桶形畸变图像对应的高分辨率图像，并输出所述高分辨率图像。
另一方面，本发明提供了一种桶形畸变图像的矫正重构装置，所述装置包括：
第一畸变恢复模块，用于接收用户输入的低分辨率桶形畸变图像，并对所述低分辨率桶形畸变图像执行畸变恢复操作，生成对应的低分辨率图像；
训练数据获取模块，用于提取所述低分辨率图像的图像特征，并在预先训练好的低分辨率稀疏字典中查找所述低分辨率图像的每个图像特征对应的字典原子，在预先训练好的多个投影矩阵中获取所述字典原子对应的投影矩阵；
第一重构模块，用于根据所述低分辨率图像的图像特征和所述投影矩阵，重构所述低分辨率桶形畸变图像对应的第一高分辨率图像特征；
第二重构模块，用于根据所述第一高分辨率图像特征和预先训练好的多个系数矩阵，重构所述低分辨率桶形畸变图像对应的第二高分辨图像特征；以及
图像输出模块，用于根据所述第二高分辨率图像特征，生成所述低分辨率桶形畸变图像对应的高分辨率图像，并输出所述高分辨率图像。
本发明先对用户输入的低分辨率桶形畸变图像进行畸变恢复，生成低分辨率图像，提取该低分辨率图像的图像特征，并在预先训练好的低分辨率稀疏字
典中查找该低分辨率图像的每个图像特征对应的字典原子，在预先训练好的多个投影矩阵中查找该字典原子对应的投影矩阵，接着，根据低分辨率图像的图像特征和投影矩阵，对低分辨率桶形畸变图像进行第一次重构，生成第一高分辨率图像特征，再根据预先训练好的多个系数矩阵进行第二次重构，生成第二高分辨率图像特征，最后得到低分辨率桶形畸变图像矫正重构后的高分辨率图像，从而通过第一次重构恢复桶形畸变图像矫正后的图像细节，进而通过第二次重构进一步地提高图像的分辨率，有效地提高了桶形畸变图像矫正重构的效率和图像质量。
图1是本发明实施例一提供的桶形畸变图像的矫正重构方法的实现流程图；
图2是本发明实施例二提供的桶形畸变图像的矫正重构方法中训练过程的实现流程；
图3是本发明实施例二提供的桶形畸变图像的矫正重构方法中非线性畸变模型中映射关系的示例图；
图4是本发明实施例二提供的桶形畸变图像的矫正重构方法中非线性畸变模型中映射关系对应高分辨率桶形畸变训练图像的示例图；
图5是本发明实施例三提供的桶形畸变图像的矫正重构装置的结构示意图；以及
图6是本发明实施例四提供的桶形畸变图像的矫正重构装置中进行训练操作的模块结构示意图。
为了使本发明的目的、技术方案及优点更加清楚明白，以下结合附图及实施例，对本发明进行进一步详细说明。应当理解，此处所描述的具体实施例仅仅用以解释本发明，并不用于限定本发明。
以下结合具体实施例对本发明的具体实现进行详细描述：
实施例一：
图1示出了本发明实施例一提供的桶形畸变图像的矫正重构方法的实现流程，为了便于说明，仅示出了与本发明实施例相关的部分，详述如下：
在步骤S101中，接收用户输入的低分辨率桶形畸变图像，并对低分辨率桶形畸变图像执行畸变恢复操作，生成对应的低分辨率图像。
在本发明实施例中，当接收到用户输入的分辨率较低的桶形畸变图像时，可对该桶形畸变图像进行畸变恢复(或矫正)，以生成对应的低分辨率图像。
具体地，在将低分辨率桶形畸变图像恢复为低分辨率图像的过程中，可先建立低分辨率桶形畸变图像的极坐标系，将极坐标系中的极坐标半径按照训练中得到的映射关系转换为新极坐标半径，并将该新极坐标半径设置为畸变恢复后图像的极坐标半径，进而生成畸变恢复后的图像，即低分辨率图像。
在步骤S102中，提取低分辨率图像的图像特征，并在预先训练好的低分辨率稀疏字典中查找低分辨率图像的每个图像特征对应的字典原子，在预先训练好的多个投影矩阵中获取字典原子对应的投影矩阵。
在本发明实施例，低分辨率稀疏字典和多个投影矩阵都是在训练过程中训练好的数据，并且在低分辨率稀疏字典中每个字典原子都对应着一个投影矩阵。具体地，低分辨率图像的图像特征对应的字典原子为低分辨率字典中离该图像特征距离最近的字典原子。
在具体实施过程中，当提取低分辨率图像的图像特征时，可先对低分辨率图像进行滤波处理，每个低分辨率图像经滤波处理后可得到多个滤波图像，接着，采用网格化方法提取这些滤波图像的图像碎片，将这些滤波图像同一位置的图像碎片进行组合，便可获得低分辨率图像的图像特征。
在步骤S103中，根据低分辨率图像的图像特征和投影矩阵，重构低分辨率桶形畸变图像对应的第一高分辨率图像特征。
在本发明实施例中，为了便于描述，将此次重构称为第一次重构。第一高
分辨率图像特征为第一次重构后得到的高分辨率图像的图像特征。具体地，可采用以下计算公式进行第一次重构：
yT,j＝PkxT,j，其中，yT,j为第j个第一高分辨率图像特征，xT,j为第j个低分辨率图像的图像特征，Pk为第k个字典原子对应的投影矩阵，第k个字典原子为低分辨率稀疏字典中离第j个低分辨率图像的图像特征距离最近的字典原子。
在步骤S104中，根据第一高分辨率图像特征和预先训练好的多个系数矩阵，重构低分辨率桶形畸变图像对应的第二高分辨率图像特征。
在本发明实施例中，为了便于描述，将此次重构称为第二次重构，由于桶形畸变图像的边缘部分和中间部分在畸变恢复过程中的信息损失程度存在差异，而在步骤S103中进行的第一次重构并未考虑到桶形畸变图像的这点图像特性，因此还需要进行一次后处理，即进行本步骤的第二次重构。第二高分辨率图像特征为第二次重构获得的第二高分辨率图像的图像特征。具体地，在第二次重构过程中，可对第一高分辨率图像特征进行分类，对于分类后的每类图像特征，采用该类图像特征对应的系数矩阵进行第二次重构，具体地，第二次重构的计算公式可为：
yR,c＝HcyT,c，其中，yR,c表示当第一高分辨率特征为第c类图像特征时重构得到的第二高分辨率图像特征，yT,c表示为第c类图像特征的第一高分辨率特征，Hc为第c类图像特征对应的系数矩阵。
优选地，可根据第一高分辨率图像特征到第一高分辨率图像中心的距离，对第一高分辨图像特征进行分类，从而基于桶形畸变图形的图形特性，将图像不同区域的特征按照不同的系数矩阵进行第二次重构，有效地提高了第二次重构后的图像分辨率和图像质量。
在步骤S105中，根据第二高分辨率图像特征，生成低分辨率桶形畸变图像对应的高分辨率图像，并输出高分辨率图像。
在本发明实施例中，将所有第二高分辨率图像特征合并起来，生成矫正重构后的高分辨率图像。
在本发明实施例中，对用户输入的低分辨率桶形畸变图像进行畸变恢复、第一次重构和第二次重构，最终生成高分辨率的图像，在第一次重构中通过预先训练好的低分辨率稀疏字典、投影矩阵，恢复了畸变矫正后图像的细节，在第二次重构中通过预先训练好的系数矩阵，对恢复细节后图像的不同区域进行对应的后处理，进一步地提高图像的分辨率和质量，解决桶形畸变图像的边缘部分和中间部分在畸变恢复过程中信息损失程度存在差异的问题，从而有效地提高桶形畸变图像矫正重构的效率和矫正重构后图像的分辨率和质量。
实施例二：
图2示出了本发明实施例二提供的桶形畸变图像的矫正重构方法中训练过程的实现流程，详述如下：
在步骤S201中，对预先存储的多张高分辨率训练图像进行处理，获得用于训练低分辨率稀疏字典、投影矩阵以及系数矩阵的低分辨率桶形畸变训练图像。
在本发明实施例中，将现有的多张高分辨率的图像作为训练过程中的训练图像，为了表达清楚、易于区分，这里将这些训练图像称为高分辨率训练图像，同样地，低分辨率桶形畸变训练图像为高分辨率训练图像经处理后得到的分辨率较低的桶形畸变图像，以用于训练低分辨率稀疏字典、投影矩阵、系数矩阵。
具体地，高分辨率训练图像经处理得到低分辨率桶形畸变处理图像可通过以下步骤来实现：
(1)根据预设的非线性畸变模型，对高分辨率训练图像进行畸变，生成高分辨率桶形畸变训练图像。
在本发明实施例中，可先为高分辨训练图像建立极坐标系，获得该极坐标系中的极坐标半径，再根据非线性畸变模型的公式，得到高分辨率桶形畸变训练图像的极坐标半径，进而生成高分辨率桶形畸变训练图像。其中，非线性畸变模型的公式为
其中，r为高分辨率训练图像的极坐标半径，rnew为高分辨率畸变训练图像的极坐标半径，ai和n为预设参数。
作为示例地，当n为2时，非线性畸变模型的公式为rnew＝a1r+a2r2，此时的运算较简便，在图3中示出了n为2、a1为1时，采用不同数值的a2进行畸变时，两个极坐标半径之间的映射关系，其中，图3坐标中的横坐标表示非畸变图像(在这里，可认为是高分辨率训练图像)的极坐标半径值，纵坐标表示畸变图像(在这里，可认为是高分辨率桶形畸变图像)的极坐标半径值，rmax为非畸变图像极坐标半径的最大值。在图4中示出了采用图3中不同的映射关系对应生成的高分辨率桶形畸变训练图像。
(2)对高分辨率桶形畸变训练图像进行处理，生成低分辨率桶形畸变训练图像。
在本发明实施例中，在对高分辨率桶形畸变训练图像进行处理时，可对高分辨率桶形畸变训练图像进行降采样，并对降采样后得到的图像进行插值处理。
优选地，在插值处理过程选用现有的bicubic插值(双三次插值)方法进行插值处理，从而在一定程度上提高低分辨率畸变训练图像的图像质量。
在步骤S202中，对低分辨率桶形畸变训练图像执行畸变恢复操作，生成相应的低分辨率训练图像。
在本发明实施例中，将低分辨率桶形畸变训练图像恢复(或矫正)为低分辨率训练图像。
具体地，对低分辨率桶形畸变训练图像执行畸变恢复操作，生成相应的低分辨率训练图像可通过下述步骤实现：
(1)根据非线性畸变模型，获取低分辨率桶形畸变训练图像的极坐标半径与低分辨率训练图像的极坐标半径之间的映射关系。
在本发明实施例中，在非线性畸变模型中可得到畸变图像与非畸变图像两者极坐标半径之间的映射关系，可以通过该映射关系，将非畸变图像畸变为畸变图像，也可将畸变图像矫正为非畸变图像，因此可根据非线性畸变模型，获得低分辨率桶形畸变训练图像的极坐标半径与低分辨率训练图像的极坐标半径之间的映射关系。
作为示例地，当非线性畸变模型中的n为2时，两极坐标半径的映射关系为
其中r1为低分辨率桶形畸变训练图像的极坐标半径，r2为低分辨率训练图像的极坐标半径。
(2)根据映射关系，将低分辨率桶形畸变训练图像转换为低分辨率训练图像。
在本发明实施例中，获得低分辨率训练图像的极坐标半径后，可根据该极坐标半径生成低分辨率训练图像。
在步骤S203中，提取低分辨率训练图像的图像特征，并提取高分辨率训练图像的图像特征。
具体地，可使用预设数量个高通滤波器对所有低分辨率训练图像进行过滤，每个低分辨率训练图像可得到相应数量个过滤后的图像，通过网格化的方法，可从这些图像中提取多个图像碎片，将这些图像中相同位置的图像碎片进行组合，可得到这些图像每个位置的图像特征。接着，可直接采用网格化的方法在高分辨率训练图像的相同位置提取图像特征，以得到高分辨率训练图像的图像特征，可见，低分辨率训练图像的图像特征与高分辨率训练图像的图像特征之间存在对应关系。
在步骤S204中，根据低分辨率训练图像的图像特征和高分辨率训练图像的图像特征，计算低分辨率稀疏字典和投影矩阵，同时根据低分辨率训练图像的图像特征、高分辨率训练图像的图像特征以及低分辨率稀疏字典，重构低分辨率桶形畸变训练图像对应的高分辨率重构图像特征。
具体地，根据低分辨率训练图像的图像特征和高分辨率训练图像的图像特征，计算低分辨率稀疏字典和投影矩阵时可通过下述步骤实现：
(1)根据低分辨率训练图像的图像特征，计算低分辨率稀疏字典，并计算低分辨率稀疏字典中每个字典原子的第一图像特征邻域和第二图像特征邻域，第一图像特征邻域由低分辨率训练图像的图像特征构成，第二图像特征邻域由高分辨率训练图像的图像特征构成。
具体地，计算低分辨率稀疏字典的公式可为
其中，δ为稀疏表示系数矩阵，x为低分辨率训练图像的图像特征，DL为低分辨率稀疏字典，λ为预设的权重因子，DL的表达公式可为
dL,k为低分辨率稀疏字典中第k个字典原子，K为预设常量。
优选地，采用现有的KNN算法(K最近邻算法)计算每个字典原子在低分辨率训练图像的图像特征中的最近邻特征，获得第一图像特征邻域，再根据低分辨率训练图像的图像特征与高分辨率训练图像的图像特征之间的对应关系，获得第二图像特征邻域，从而有效地提高最近邻特征的计算效率。
(2)根据每个字典原子的第一图像特征邻域和第二图像特征邻域，计算得到每个字典原子对应的投影矩阵。
具体地，投影矩阵的计算公式可为
其中，Pk为第k个字典原子对应的投影矩阵，NH,k为第二图像特征邻域，NL,k为第一图像特征邻域。
在本发明实施例中，高分辨率重构图像特征为训练图像重构后得到的高分辨率图像的图像特征，具体地，高分辨率重构图像特征的重构公式可为：
其中，yH,i为第i个高分辨率重构图像特征。
在步骤S205中，根据高分辨率重构图像特征和高分辨率训练图像的图像特征，计算系数矩阵。
在本发明实施例中，可分别对高分辨率重构图像和高分辨率训练图像的图像特征进行分类，并计算分类后的每类图像特征对应的系数矩阵，具体地，系数矩阵的计算公式可为
其中，Hc为第c类图像特征对应的系数矩阵，yH,c为第c类图像特征的高分辨率重构图像特征，yc为第c类图像特征的高分辨率训练图像的图像特征。
在本发明实施例中，通过预设的非线性畸变模型，对高分辨率训练图像进
行畸变，得到低分辨率桶形畸变训练图像，再对该低分辨率桶形畸变图像畸形进行畸变恢复、重构等操作，训练得到低分辨率稀疏字典、每个字典元素对应的投影矩阵以及每类图像特征对应的系数矩阵，从而有效地提高实施例一中对低分辨率桶形畸变图像进行矫正重构的效果。
本领域普通技术人员可以理解实现上述实施例方法中的全部或部分步骤是可以通过程序来指令相关的硬件来完成，所述的程序可以存储于一计算机可读取存储介质中，所述的存储介质，如ROM/RAM、磁盘、光盘等。
实施例三：
图5示出了本发明实施例三提供的桶形畸变图像的矫正重构装置的结构，为了便于说明，仅示出了与本发明实施例相关的部分，其中包括：
第一畸变恢复模块51，用于接收用户输入的低分辨率桶形畸变图像，并对低分辨率桶形畸变图像执行畸变恢复操作，生成对应的低分辨率图像；
训练数据获取模块52，用于提取低分辨率图像的图像特征，并在预先训练好的低分辨率稀疏字典中查找低分辨率图像的每个图像特征对应的字典原子，在预先训练好的多个投影矩阵中获取字典原子对应的投影矩阵；
第一重构模块53，用于根据低分辨率图像的图像特征和投影矩阵，重构低分辨率桶形畸变图像对应的第一高分辨率图像特征；
第二重构模块54，用于根据第一高分辨率图像特征和预先训练好的多个系数矩阵，重构低分辨率桶形畸变图像对应的第二高分辨图像特征；以及
图像输出模块55，用于根据第二高分辨率图像特征，生成低分辨率桶形畸变图像对应的高分辨率图像，并输出高分辨率图像。
在本发明实施例中，对用户输入的低分辨率桶形畸变图像进行畸变恢复、第一次重构和第二次重构，最终生成高分辨率的图像，在第一次重构中通过预先训练好的低分辨率稀疏字典、投影矩阵，恢复了畸变矫正后图像的细节，在第二次重构中通过预先训练好的系数矩阵，对恢复细节后图像的不同区域进行对应的后处理，进一步地提高图像的分辨率和质量，解决桶形畸变图像的边缘
部分和中间部分在畸变恢复过程中信息损失程度存在差异的问题，从而有效地提高桶形畸变图像矫正重构的效率和矫正重构后图像的分辨率和质量。本发明实施例中各模块的具体实施方式可参考前述实施例一中各步骤的描述，在此不再赘述。
实施例四：
图6示出了本发明实施例四提供的桶形畸变图像的矫正重构装置中进行训练操作的模块结构，为了便于说明，仅示出了与本发明实施例相关的部分，其中包括：
训练图像处理模块61，用于对预先存储的多张高分辨率训练图像进行处理，获得用于训练低分辨率稀疏字典、投影矩阵以及系数矩阵的低分辨率桶形畸变训练图像；
第二畸变恢复模块62，用于对低分辨率桶形畸变训练图像执行畸变恢复操作，生成相应的低分辨率训练图像；
特征提取模块63，用于提取低分辨率训练图像的图像特征，并提取高分辨率训练图像的图像特征；
训练重构模块64，用于根据低分辨率训练图像的图像特征和高分辨率训练图像的图像特征，计算低分辨率稀疏字典和投影矩阵，同时根据低分辨率训练图像的图像特征、高分辨率训练图像的图像特征以及低分辨率稀疏字典，重构低分辨率桶形畸变训练图像对应的高分辨率重构图像特征；以及
系数矩阵计算模块65，用于根据高分辨率重构图像特征和高分辨率训练图像的图像特征，计算系数矩阵。
因此，优选地，训练图像处理模块61还包括图像畸变模块611和图像分辨率处理模块612，其中：
图像畸变模块611，用于根据预设的非线性畸变模型，对高分辨率训练图像进行畸变，生成高分辨率桶形畸变训练图像；以及
图像分辨率处理模块612，用于对高分辨率桶形畸变训练图像进行处理，
生成低分辨率桶形畸变训练图像。
优选地，第二畸变恢复模块62还包括映射关系获取模块621和图像转换模块622，其中：
映射关系获取模块621，用于根据非线性畸变模型，获取低分辨率桶形畸变训练图像的极坐标半径与低分辨率训练图像的极坐标半径之间的映射关系；以及
图像转换模块622，用于根据映射关系，将低分辨率桶形畸变训练图像转换为低分辨率训练图像。
在本发明实施例中，通过预设的非线性畸变模型，对高分辨率训练图像进行畸变，得到低分辨率桶形畸变训练图像，再对该低分辨率桶形畸变图像畸形进行畸变恢复、第一次重构等操作，训练得到低分辨率稀疏字典、每个字典元素对应的投影矩阵以及每类图像特征对应的系数矩阵，从而有效地提高实施例三中对低分辨率桶形畸变图像进行矫正重构的效果。
在本发明实施例中，桶形畸变图像的矫正重构装置的各单元可由相应的硬件或软件单元实现，各单元可以为独立的软、硬件单元，也可以集成为一个软、硬件单元，在此不用以限制本发明。各单元的具体实施方式可参考前述实施例二中各步骤的描述，在此不再赘述。
以上所述仅为本发明的较佳实施例而已，并不用以限制本发明，凡在本发明的精神和原则之内所作的任何修改、等同替换和改进等，均应包含在本发明的保护范围之内。
Claims (10)
- 一种桶形畸变图像的矫正重构方法，其特征在于，所述方法包括下述步骤：接收用户输入的低分辨率桶形畸变图像，并对所述低分辨率桶形畸变图像执行畸变恢复操作，生成对应的低分辨率图像；提取所述低分辨率图像的图像特征，并在预先训练好的低分辨率稀疏字典中查找所述低分辨率图像的每个图像特征对应的字典原子，在预先训练好的多个投影矩阵中获取所述字典原子对应的投影矩阵；根据所述低分辨率图像的图像特征和所述投影矩阵，重构所述低分辨率桶形畸变图像对应的第一高分辨率图像特征；根据所述第一高分辨率图像特征和预先训练好的多个系数矩阵，重构所述低分辨率桶形畸变图像对应的第二高分辨率图像特征；根据所述第二高分辨率图像特征，生成所述低分辨率桶形畸变图像对应的高分辨率图像，并输出所述高分辨率图像。
- 如权利要求1所述的方法，其特征在于，接收用户输入的低分辨率桶形畸变图像的步骤之前，所述方法还包括：对预先存储的多张高分辨率训练图像进行处理，获得用于训练所述低分辨率稀疏字典、投影矩阵以及系数矩阵的低分辨率桶形畸变训练图像；对所述低分辨率桶形畸变训练图像执行所述畸变恢复操作，生成相应的低分辨率训练图像；提取所述低分辨率训练图像的图像特征，并提取所述高分辨率训练图像的图像特征；根据所述低分辨率训练图像的图像特征和所述高分辨率训练图像的图像特征，计算所述低分辨率稀疏字典和所述投影矩阵，同时根据所述低分辨率训练图像的图像特征、所述高分辨率训练图像的图像特征以及所述低分辨率稀疏字典，重构所述低分辨率桶形畸变训练图像对应的高分辨率重构图像特征；根据所述高分辨率重构图像特征和所述高分辨率训练图像的图像特征，计算所述系数矩阵。
- 如权利要求2所述的方法，其特征在于，对预先存储的多张高分辨率训练图像进行处理，获得用于训练所述低分辨率稀疏字典、投影矩阵以及系数矩阵的低分辨率桶形畸变训练图像的步骤，包括：根据预设的非线性畸变模型，对所述高分辨率训练图像进行畸变，生成高分辨率桶形畸变训练图像；对所述高分辨率桶形畸变训练图像进行处理，生成所述低分辨率桶形畸变训练图像；所述非线性畸变模型的公式为其中，r为所述高分辨率训练图像的极坐标半径，rnew为所述高分辨率桶形畸变训练图像的极坐标半径，ai和n为预设参数。
- 如权利要求3所述的方法，其特征在于，对所述低分辨率桶形畸变训练图像执行所述畸变恢复操作，生成相应的低分辨率训练图像的步骤，包括：根据所述非线性畸变模型，获取所述低分辨率桶形畸变训练图像的极坐标半径与所述低分辨率训练图像的极坐标半径之间的映射关系；根据所述映射关系，将所述低分辨率桶形畸变训练图像转换为所述低分辨率训练图像。
- 如权利要求2所述的方法，其特征在于，根据所述低分辨率训练图像的图像特征和所述高分辨率训练图像的图像特征，计算所述低分辨率稀疏字典和所述投影矩阵的步骤，包括：根据所述低分辨率训练图像的图像特征，计算所述低分辨率稀疏字典，并计算所述低分辨率稀疏字典中每个字典原子的第一图像特征邻域和第二图像特征邻域，所述第一图像特征邻域由所述低分辨率训练图像的图像特征构成，所述第二图像特征邻域由所述高分辨率训练图像的图像特征构成；根据所述每个字典原子的第一图像特征邻域和第二图像特征邻域，计算所述每个字典原子对应的投影矩阵。
- 如权利要求2所述的方法，其特征在于，根据所述高分辨率重构图像特征和所述高分辨率训练图像的图像特征，计算所述系数矩阵的步骤，包括：对所述高分辨率重构图像特征、所述高分辨率训练图像的图像特征分别进行分类，并计算分类后的每类图像特征对应的系数矩阵。
- 一种桶形畸变图像的矫正重构装置，其特征在于，所述装置包括：第一畸变恢复模块，用于接收用户输入的低分辨率桶形畸变图像，并对所述低分辨率桶形畸变图像执行畸变恢复操作，生成对应的低分辨率图像；训练数据获取模块，用于提取所述低分辨率图像的图像特征，并在预先训练好的低分辨率稀疏字典中查找所述低分辨率图像的每个图像特征对应的字典原子，在预先训练好的多个投影矩阵中获取所述字典原子对应的投影矩阵；第一重构模块，用于根据所述低分辨率图像的图像特征和所述投影矩阵，重构所述低分辨率桶形畸变图像对应的第一高分辨率图像特征；第二重构模块，用于根据所述第一高分辨率图像特征和预先训练好的多个系数矩阵，重构所述低分辨率桶形畸变图像对应的第二高分辨图像特征；以及图像输出模块，用于根据所述第二高分辨率图像特征，生成所述低分辨率桶形畸变图像对应的高分辨率图像，并输出所述高分辨率图像。
- 如权利要求7所述的装置，其特征在于，所述装置还包括：训练图像处理模块，用于对预先存储的多张高分辨率训练图像进行处理，获得用于训练所述低分辨率稀疏字典、投影矩阵以及系数矩阵的低分辨率桶形畸变训练图像；第二畸变恢复模块，用于对所述低分辨率桶形畸变训练图像执行所述畸变恢复操作，生成相应的低分辨率训练图像；特征提取模块，用于提取所述低分辨率训练图像的图像特征，并提取所述高分辨率训练图像的图像特征；训练重构模块，用于根据所述低分辨率训练图像的图像特征和所述高分辨率训练图像的图像特征，计算所述低分辨率稀疏字典和所述投影矩阵，同时根据所述低分辨率训练图像的图像特征、所述高分辨率训练图像的图像特征以及所述低分辨率稀疏字典，重构所述低分辨率桶形畸变训练图像对应的高分辨率重构图像特征；以及系数矩阵计算模块，用于根据所述高分辨率重构图像特征和所述高分辨率训练图像的图像特征，计算所述系数矩阵。
- 如权利要求8所述的装置，其特征在于，所述训练图像处理模块包括：图像畸变模块，用于根据预设的非线性畸变模型，对所述高分辨率训练图像进行畸变，生成高分辨率桶形畸变训练图像；以及图像分辨率处理模块，用于对所述高分辨率桶形畸变训练图像进行处理，生成所述低分辨率桶形畸变训练图像。
- 如权利要求9所述的装置，其特征在于，所述第二畸变恢复模块包括：映射关系获取模块，用于根据所述非线性畸变模型，获取所述低分辨率畸变训练图像的极坐标半径与所述低分辨率训练图像的极坐标半径之间的映射关系；以及图像转换模块，用于根据所述映射关系，将所述低分辨率畸变训练图像转换为所述低分辨率训练图像。
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| PCT/CN2016/112068 WO2018119565A1 (zh) | 2016-12-26 | 2016-12-26 | 一种桶形畸变图像的矫正重构方法及装置 |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| PCT/CN2016/112068 WO2018119565A1 (zh) | 2016-12-26 | 2016-12-26 | 一种桶形畸变图像的矫正重构方法及装置 |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| WO2018119565A1 true WO2018119565A1 (zh) | 2018-07-05 |
Family
ID=62706981
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| PCT/CN2016/112068 WO2018119565A1 (zh) | 2016-12-26 | 2016-12-26 | 一种桶形畸变图像的矫正重构方法及装置 |
Country Status (1)
| Country | Link |
|---|---|
| WO (1) | WO2018119565A1 (zh) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109635860A (zh) * | 2018-12-04 | 2019-04-16 | 科大讯飞股份有限公司 | 图像分类方法及系统 |
| CN111784580A (zh) * | 2020-06-30 | 2020-10-16 | 北京金山云网络技术有限公司 | 图像的超分辨率方法、装置和服务器 |
| CN113538283A (zh) * | 2021-07-22 | 2021-10-22 | 浙江赫千电子科技有限公司 | 一种冗余鱼眼摄像头拍摄图像的畸变矫正方法 |
| CN113538283B (zh) * | 2021-07-22 | 2024-04-30 | 浙江赫千电子科技有限公司 | 一种冗余鱼眼摄像头拍摄图像的畸变矫正方法 |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101950365A (zh) * | 2010-08-30 | 2011-01-19 | 西安电子科技大学 | 基于ksvd字典学习的多任务超分辨率图像重构方法 |
| CN102156875A (zh) * | 2011-03-25 | 2011-08-17 | 西安电子科技大学 | 基于多任务ksvd字典学习的图像超分辨率重构方法 |
| CN103530863A (zh) * | 2013-10-30 | 2014-01-22 | 广东威创视讯科技股份有限公司 | 一种多级重构的图像超分辨率方法 |
| US20140043477A1 (en) * | 2012-08-13 | 2014-02-13 | Politechnika Poznanska | Method for processing wide angle images with barrel distortion and a surveillance system |
-
2016
- 2016-12-26 WO PCT/CN2016/112068 patent/WO2018119565A1/zh active Application Filing
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101950365A (zh) * | 2010-08-30 | 2011-01-19 | 西安电子科技大学 | 基于ksvd字典学习的多任务超分辨率图像重构方法 |
| CN102156875A (zh) * | 2011-03-25 | 2011-08-17 | 西安电子科技大学 | 基于多任务ksvd字典学习的图像超分辨率重构方法 |
| US20140043477A1 (en) * | 2012-08-13 | 2014-02-13 | Politechnika Poznanska | Method for processing wide angle images with barrel distortion and a surveillance system |
| CN103530863A (zh) * | 2013-10-30 | 2014-01-22 | 广东威创视讯科技股份有限公司 | 一种多级重构的图像超分辨率方法 |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109635860A (zh) * | 2018-12-04 | 2019-04-16 | 科大讯飞股份有限公司 | 图像分类方法及系统 |
| CN109635860B (zh) * | 2018-12-04 | 2023-04-07 | 科大讯飞股份有限公司 | 图像分类方法及系统 |
| CN111784580A (zh) * | 2020-06-30 | 2020-10-16 | 北京金山云网络技术有限公司 | 图像的超分辨率方法、装置和服务器 |
| CN113538283A (zh) * | 2021-07-22 | 2021-10-22 | 浙江赫千电子科技有限公司 | 一种冗余鱼眼摄像头拍摄图像的畸变矫正方法 |
| CN113538283B (zh) * | 2021-07-22 | 2024-04-30 | 浙江赫千电子科技有限公司 | 一种冗余鱼眼摄像头拍摄图像的畸变矫正方法 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Anwar et al. | Densely residual laplacian super-resolution | |
| CN111028150B (zh) | 一种快速时空残差注意力视频超分辨率重建方法 | |
| Zhang et al. | Image super-resolution via dual-dictionary learning and sparse representation | |
| CN107392852B (zh) | 深度图像的超分辨率重建方法、装置、设备及存储介质 | |
| WO2017004890A1 (zh) | 基于迭代投影重建的字典类图像超分辨率系统及方法 | |
| Sun et al. | Lightweight image super-resolution via weighted multi-scale residual network | |
| WO2015180053A1 (zh) | 一种快速超分辨率图像重建方法和装置 | |
| CN111369466B (zh) | 基于可变形卷积的卷积神经网络的图像畸变矫正增强方法 | |
| CN107680043B (zh) | 基于图模型的单幅图像超分辨率输出方法 | |
| CN111640059B (zh) | 基于高斯混合模型的多字典图像超分辨方法 | |
| CN111553867B (zh) | 一种图像去模糊方法、装置、计算机设备及存储介质 | |
| CN109636722B (zh) | 一种基于稀疏表示的在线字典学习超分辨率重建的方法 | |
| CN103020898A (zh) | 序列虹膜图像超分辨率重建方法 | |
| CN104899835A (zh) | 基于盲模糊估计与锚定空间映射的图像超分辨处理方法 | |
| Liu et al. | Multi-scale residual hierarchical dense networks for single image super-resolution | |
| WO2018119565A1 (zh) | 一种桶形畸变图像的矫正重构方法及装置 | |
| Zhang et al. | Deep spatial adaptive network for real image demosaicing | |
| KR20120088350A (ko) | 고해상도 영상 생성 장치 | |
| Chen et al. | Guided dual networks for single image super-resolution | |
| CN117315735A (zh) | 基于先验信息与注意力机制的人脸超分辨率重建方法 | |
| Wang et al. | Heatmap-aware pyramid face hallucination | |
| Sun et al. | Distilling with residual network for single image super resolution | |
| Bareja et al. | An improved iterative back projection based single image super resolution approach | |
| Qin et al. | Remote sensing image super-resolution using multi-scale convolutional neural network | |
| CN104123707A (zh) | 一种基于局部秩先验的单幅图像超分辨率重建方法 |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| 121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 16925469 Country of ref document: EP Kind code of ref document: A1 |
|
| NENP | Non-entry into the national phase |
Ref country code: DE |
|
| 32PN | Ep: public notification in the ep bulletin as address of the adressee cannot be established |
Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 16.09.2019) |
|
| 122 | Ep: pct application non-entry in european phase |
Ref document number: 16925469 Country of ref document: EP Kind code of ref document: A1 |