WO2025200534A1 - 基于改进特征描述符和knn搜索的眼底图像特征点匹配方法 - Google Patents
基于改进特征描述符和knn搜索的眼底图像特征点匹配方法Info
- Publication number
- WO2025200534A1 WO2025200534A1 PCT/CN2024/135993 CN2024135993W WO2025200534A1 WO 2025200534 A1 WO2025200534 A1 WO 2025200534A1 CN 2024135993 W CN2024135993 W CN 2024135993W WO 2025200534 A1 WO2025200534 A1 WO 2025200534A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- image
- feature
- point
- matching
- fundus
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/18—Eye characteristics, e.g. of the iris
- G06V40/197—Matching; Classification
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/40—Image enhancement or restoration using histogram techniques
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/74—Image or video pattern matching; Proximity measures in feature spaces
- G06V10/75—Organisation of the matching processes, e.g. simultaneous or sequential comparisons of image or video features; Coarse-fine approaches, e.g. multi-scale approaches; using context analysis; Selection of dictionaries
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30004—Biomedical image processing
- G06T2207/30041—Eye; Retina; Ophthalmic
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Definitions
- the present invention relates to the field of computer vision, and in particular to a fundus image feature point matching method based on improved feature descriptors and KNN search.
- Feature recognition and stitching and fusion of fundus images are crucial components of many medical research fields and advanced artificial intelligence (AI) technology development.
- fundus health screening is a key indicator of human health screening. Because most medical imaging cameras have limited shooting angles, multiple fundus images taken by a single instrument must be stitched and fused to form a panoramic fundus image before medical screening. This panoramic image is then used for case screening and health testing.
- accuracy of human organ recognition and detection such as eye and lens recognition, is extremely high. This requires that recognition instruments be able to perform detailed and accurate characterization of the eye before recognition. Therefore, accurate feature extraction and image stitching are also required within the device. Feature extraction and matching of fundus images play a crucial role in these applications.
- the feature extraction and matching of fundus images includes three steps: image preprocessing, image feature extraction, and image feature matching.
- Preprocessing often uses corresponding image enhancement operations to highlight image features based on different image characteristics.
- Feature extraction mainly includes classic local feature extraction methods such as SIFT, SURF, ORB, and KAZE.
- the FAST detector is added to detect feature points and the MLD feature description method is added to improve detection accuracy and facilitate feature matching.
- deep learning methods such as neural networks and genetic algorithms are also added.
- the present invention is proposed in view of the problems existing in the existing fundus image feature point matching method based on improved feature descriptors and KNN search. Therefore, the problem to be solved by the present invention is how to provide a fundus image feature point matching method based on improved feature descriptors and KNN search.
- the present invention provides the following technical solutions:
- An improved image scale space pyramid is constructed, and the Gaussian difference pyramid technique is used to compare the pixel values of the fundus representation image to preliminarily determine the extreme points, locate the position and direction of the key pixels, and use the improved feature vector to generate feature descriptors;
- the KNN algorithm is used to search for the nearest neighbor and next nearest neighbor matching points of the key points in each group of fundus images, eliminate erroneous matching, and retain excellent matching pairs.
- the image preprocessing includes the following steps:
- the grayscale values and spatial domain values of the pixels around the center of each pixel in the evenly distributed fundus image histogram are calculated according to the grayscale value transformation table.
- the pixel grayscale values and spatial domain values are weighted averaged and bilateral filtering is performed to enhance the image and obtain the fundus representation image.
- the generating of the uniformly distributed fundus image histogram includes the following steps:
- n represents the number of pixels in an image
- l represents the total number of gray levels
- nk represents the number of pixels with gray level rk ;
- mapping function is calculated based on the histogram gray value uniformity principle and the cumulative distribution function.
- the calculation formula is as follows:
- the mapped value of each pixel is calculated according to the mapping function to generate a uniformly distributed fundus image histogram.
- the specific steps of the bilateral filtering enhancement include:
- Wd represents the point distance weight
- Wi represents the point value weight
- is the Euclidean distance between point p and point q
- is the pixel difference between the two points, represents the pixel variance, represents the distance variance
- the fusion weight of each point on the image is calculated based on the point distance weight and the point value weight.
- the calculation formula is as follows:
- Wp represents the weight after fusion
- A represents the calculated image pixel set
- the filtered pixel intensity is calculated by taking the weighted average of the pixels surrounding each pixel.
- the calculation formula is as follows:
- I'p represents the pixel intensity after filtering.
- the specific steps of generating feature descriptors include:
- a Gaussian pyramid is constructed based on fundus representation images. The first three layers of each image pyramid are taken to calculate the difference image, generating a Gaussian difference pyramid. The improved scale space image is obtained by convolving the Gaussian difference pyramid with the original image. The pixel values of the pixels in the area of the pyramid are compared to determine the extreme points.
- a directional reference is assigned to each local feature of the image.
- the value and direction of the key pixel are calculated by combining the scale-direction distribution characteristics of adjacent pixels.
- the calculation formula of the value and direction of the key pixel is as follows:
- I (x,y) represents the value of the extreme point, Indicates the direction of the extreme point, (x+1, y+1) and (x-1, y-1) represent the information representation of the adjacent pixels of the extreme point in the scale space;
- An image scale space displaying comprehensive information under multi-scale parameters is constructed by convolving a two-dimensional Gaussian function with the original fundus image.
- the image is downsampled using the Gaussian function to generate an image pyramid model.
- the error elimination matching includes the following steps:
- R represents the proportional threshold
- N represents the total number of feature points
- wk represents the weight of the kth feature point
- d( xk , yk ) represents the Euclidean distance between the kth feature point and its nearest neighbor
- ⁇ represents the normalization coefficient
- wj represents the weight coefficient for normalizing the weights of all feature points.
- the image preprocessing module is used to obtain fundus images, convert them into grayscale images, perform histogram equalization and bilateral filtering enhancement, and obtain fundus representation images;
- the feature point detection module is used to obtain improved scale space and Gaussian difference pyramid technology, compare pixel values to determine key pixels, and calculate the position and direction information of the points;
- the feature descriptor generation module is used to divide the sub-region around the key point, calculate the feature vector of each sub-region, and generate the feature descriptor of the key point by combining the distance weight superposition;
- the feature matching module is used to find the nearest neighbor matching point and the next nearest neighbor matching point of each key point using the KNN nearest neighbor search algorithm, and calculate the distance ratio to eliminate false matches.
- 1 and 2 which are the first embodiment of the present invention, provide a method for matching feature points of fundus images based on an improved feature descriptor and KNN search.
- Input the fundus image set to be stitched perform image preprocessing on the image set, and obtain the fundus representation image.
- image preprocessing includes the following steps:
- the grayscale values and spatial domain values of the pixels around the center of each pixel in the evenly distributed fundus image histogram are calculated according to the grayscale value transformation table.
- the pixel grayscale values and spatial domain values are weighted averaged and bilateral filtering is performed to enhance the image and obtain the fundus representation image.
- generating a uniformly distributed fundus image histogram includes the following steps:
- n represents the number of pixels in an image
- l represents the total number of gray levels
- nk represents the number of pixels with gray level rk ;
- mapping function is calculated based on the histogram gray value uniformity principle and the cumulative distribution function.
- the calculation formula is as follows:
- An improved image scale space pyramid is constructed, and the Gaussian difference pyramid technology is used to compare the pixel values of the fundus representation image to preliminarily determine the extreme points, locate the position and direction of the key pixels, and use the improved feature vector to generate feature descriptors.
- I (x,y) represents the value of the extreme point, Indicates the direction of the extreme point, (x+1, y+1) and (x-1, y-1) represent the information representation of the adjacent pixels of the extreme point in the scale space;
- the sub-region around the key pixel point is divided into 4 sub-regions, and each sub-region is defined to include 16 directional features.
- the feature information of the key pixel points of the fundus image is represented for feature matching.
- An image scale space displaying comprehensive information under multi-scale parameters is constructed by convolving a two-dimensional Gaussian function with the original fundus image.
- the image is downsampled using the Gaussian function to generate an image pyramid model.
- the pixel values of the difference pyramid image are compared with the pixel values in the surrounding adjacent scale range to determine the extreme points.
- the KNN algorithm is used to search for the nearest neighbor and next nearest neighbor matching points of the key points in each group of fundus images, eliminate erroneous matching, and retain excellent matching pairs.
- eliminating error matching includes the following steps:
- R represents the proportional threshold
- N represents the total number of feature points
- wk represents the weight of the kth feature point
- d( xk , yk ) represents the Euclidean distance between the kth feature point and its nearest neighbor
- ⁇ represents the normalization coefficient
- wj represents the weight coefficient for normalizing the weights of all feature points.
- the ratio of the nearest neighbor distance to the next nearest neighbor distance is less than the proportional threshold, it means that the matching point pair meets the feature point requirements and the matching point pair is accepted;
- the ratio of the nearest neighbor distance to the next nearest neighbor distance is greater than the proportional threshold, it means that the matching point pair does not meet the feature point requirements and the matching point pair is eliminated.
- the k value in the KNN algorithm is 2.
- This embodiment also provides a computer device, which is applicable to the fundus image feature point matching method based on improved feature descriptors and KNN search, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the fundus image feature point matching method based on improved feature descriptors and KNN search proposed in the above embodiment.
- the display screen of the computer device may be a liquid crystal display or an electronic ink display screen
- the input device of the computer device may be a touch layer covering the display screen, or a button, trackball or touchpad provided on the housing of the computer device, or an external keyboard, touchpad or mouse.
- This embodiment also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the fundus image feature point matching method based on improved feature descriptors and KNN search as proposed in the above embodiment;
- the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, disk or optical disk.
- SRAM static random access memory
- EEPROM electrically erasable programmable read-only memory
- EPROM erasable programmable read-only memory
- PROM programmable read-only memory
- ROM read-only memory
- magnetic memory magnetic memory
- flash memory disk or optical disk.
- the present invention realizes efficient processing and feature extraction of fundus images, provides a reliable basis for subsequent matching, and helps to improve the accuracy and robustness of matching.
- image preprocessing By optimizing image preprocessing, the quality and characterization ability of the image are improved, providing better input for subsequent steps, helping to enhance the characteristics of the fundus characterization image, and improving the reliability of matching.
- the grayscale distribution of the image is optimized, which helps to improve the contrast and information expression ability of the image, and provides a more discriminative image representation for feature extraction.
- the bilateral filtering enhancement step the calculation of the point distance weight and the point value weight achieves a bilateral filtering effect that smoothes the image while maintaining the edge details of the image, further improving the quality of the image.
- the distance weight superposition method of the sub-region feature vector combines local and global information, making the descriptor more robust to scene changes.
- the multiple steps in the present invention cooperate with each other to significantly improve the image key point matching effect, eliminate false matching through the distance ratio judgment mechanism, and ensure the reliability of the matching result.
- FIG. 1 and 2 which illustrate a second embodiment of the present invention, provide a fundus image feature point matching method based on an improved feature descriptor and KNN search.
- scientific demonstration is performed through economic benefit calculations and simulation experiments.
- Dataset preparation Two sets of fundus image datasets are selected, one for applying the method of the present invention and the other for applying the traditional SIFT algorithm.
- Image preprocessing Grayscale conversion, histogram equalization and bilateral filtering are performed on the two sets of fundus image data sets.
- Feature point detection and description In the method of the present invention, an improved scale space pyramid is constructed, and Gaussian difference pyramid technology is used to determine the extreme points, and then an improved feature descriptor is generated. In the traditional SIFT method, a standard feature detection and description algorithm is used.
- the table data shows that the proposed method improves the accuracy of feature point matching by an average of 8.7% and the processing speed by an average of 0.7 seconds. This significant improvement is mainly attributed to the following innovations:
- Optimized feature descriptors More accurate descriptions of feature points, enhancing accuracy during the matching process.
- the KNN algorithm combined with the distance ratio judgment mechanism effectively reduces mismatches and improves the overall matching accuracy.
- the embodiments of the present invention demonstrate its advantages in matching feature points in fundus images, particularly in terms of matching accuracy and processing speed. These advantages are of great significance for fundus image analysis and improve the reliability and efficiency of the algorithm in practical applications.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Health & Medical Sciences (AREA)
- Databases & Information Systems (AREA)
- Evolutionary Computation (AREA)
- Computing Systems (AREA)
- Medical Informatics (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- Ophthalmology & Optometry (AREA)
- Human Computer Interaction (AREA)
- Eye Examination Apparatus (AREA)
Abstract
本发明公开了基于改进特征描述符和KNN搜索的眼底图像特征点匹配方法,涉及计算机视觉,包括输入需要拼接的眼底图像集,对图片集进行图像预处理,获得眼底表征图像;构建改进的图像尺度空间金字塔,并运用高斯差分金字塔技术比较眼底表征图像像素值初步判定极值点,定位关键像素点的位置与方向,并采用改进特征向量生成特征描述符;通过KNN算法搜索每组眼底图像中关键点的最近邻与次近邻匹配点,消除误差匹配,保留优秀匹配对。本发明实现了对眼底图像的高效处理和特征提取,优化了图像的灰度分布,提高了图像的对比度和信息表达能力,同时使图像关键点匹配效果显著提升,通过距离比值判断机制消除了误匹配,保证了匹配结果的可靠性。
Description
本发明涉及计算机视觉领域,特别是基于改进特征描述符和KNN搜索的眼底图像特征点匹配方法。
眼底图像的特征识别与拼接融合是许多医疗科研领域及先进人工智能科技研发的重要部分。在医疗行业,眼底健康筛查是人体健康筛查的一项重要指标,由于大多医疗检测摄像仪器的拍摄角度受限,在进行医疗筛查前需要对一个仪器拍摄的多张不同眼底图像进行拼接融合形成眼底全景图,再基于全景图进行病例筛查与健康检测;在一些先进人工智能研发领域,人们对眼球识别、晶体识别等人体器官识别检测精度要求极高,这就需要识别仪器可以在识别前对眼球进行详尽精准的表征,因此器械中也需要对眼球有准确的特征提取与图像拼接结果。眼底图像的特征提取与匹配在这些应用中起到重要作用。
由于人体眼球器官差异性大、体积小,以及拍摄仪器的精度、拍摄时光线等等问题导致拍摄出的眼底图像多存在噪声点多、色调不均衡、清晰度较低等问题,整体质量较差,因此眼底图像的特征提取与匹配难度较高,而研究相应快速、高精度的应对方法则具有重要价值。
眼底图像的特征提取与匹配工作包括图像预处理、图像特征提取与图像特征匹配三个步骤。预处理往往根据不同图像特质采用相应的图像增强操作手段对图像特征进行突出化操作,特征提取主要包括sift、surf和ORB、KAZE等经典局部特征提取方法,并在过程中加入FAST检测器检测特征点、MLD特征描述方法等提高检测精度并便于特征匹配,并随着人工智能的发展在其中加入神经网络、遗传算法等深度学习方法。虽然图像特征提取算法已有许多成熟的案例,但眼底图像由于其噪声点多、图像阴影面积大、清晰度低等问题导致上的常规图像特征提取与匹配算法应用于医疗方面效果并不优良,且目前专注于该方面研究并不充分,针对眼底图像特有的特征分布方式与特征信息所改进的特征提取与匹配方法并没有。
鉴于现有的基于改进特征描述符和KNN搜索的眼底图像特征点匹配方法存在的问题,提出了本发明。因此,本发明所要解决的问题在于如何提供基于改进特征描述符和KNN搜索的眼底图像特征点匹配方法。
为解决上述技术问题,本发明提供如下技术方案:
第一方面,本发明实施例提供了基于改进特征描述符和KNN搜索的眼底图像特征点匹配方法,其包括,输入需要拼接的眼底图像集并进行图像预处理,获得眼底表征图像;
构建改进的图像尺度空间金字塔,并运用高斯差分金字塔技术比较眼底表征图像像素值初步判定极值点,定位关键像素点的位置与方向,采用改进特征向量生成特征描述符;
通过KNN算法搜索每组眼底图像中关键点的最近邻与次近邻匹配点,消除误差匹配,保留优秀匹配对。
作为本发明所述基于改进特征描述符和KNN搜索的眼底图像特征点匹配方法的一种优选方案,其中:所述图像预处理包括以下步骤,
获取眼底图像数据集;
输入眼底图像并转化为灰度图像,求出眼底图像灰度直方图,计算灰度值变换表,根据给定眼底图像灰度直方图生成均匀分布的眼底图像直方图;
根据灰度值变换表计算均匀分布的眼底图像直方图中各像素中心周围的像素点灰度值与空域值,对像素点灰度值与空域值进行加权平均,进行双边滤波过滤增强,得到眼底表征图像。
作为本发明所述基于改进特征描述符和KNN搜索的眼底图像特征点匹配方法的一种优选方案,其中:所述生成均匀分布的眼底图像直方图包括以下步骤,
将眼底图像转化为灰度图像,并扫描原始灰度图像的每一个像素点,得出眼底图像灰度直方图;
计算眼底图像灰度直方图的累积分布函数,计算公式如下:
其中,n表示一幅图像的像素点数目,l表示灰度级总数,nk表示灰度级为rk的像素点数目;
根据直方图灰度值均匀原理和累积分布函数计算映射函数,计算公式如下:
根据映射函数计算出每个像素点经过映射后的数值,生成均匀分布的眼底图像直方图。
作为本发明所述基于改进特征描述符和KNN搜索的眼底图像特征点匹配方法的一种优选方案,其中:所述双边滤波过滤增强的具体步骤包括,
输入经过直方图均匀后的眼底图像,针对每张图像像素点计算双边滤波器中的点距权重与点值权重,计算公式如下:
其中,Wd表示点距权重,Wi表示点值权重,||p-q||是点p与点q之间的欧式距离,|Ip-Iq|是两点的像素差值,表示像素方差,表示距离方差;
根据点距权重与点值权重计算图像上每个点的融合权重,计算公式如下:
其中,Wp表示融合之后的权重,A表示计算的图像像素点集;
通过每个像素点周边像素的加权平均值来计算过滤后的像素强度,计算公式如下:
其中,I'p表示过滤后的像素强度。
作为本发明所述基于改进特征描述符和KNN搜索的眼底图像特征点匹配方法的一种优选方案,其中:所述生成特征描述符的具体步骤包括,
根据眼底表征图像构建高斯金字塔,取每组图像金字塔的前三层计算差分图像,生成高斯差分金字塔,通过高斯差分金字塔与原图像卷积得到改进尺度空间图像,比较金字塔中区域内像素点的像素值判断极值点;
根据极值点的尺度不变性,分配方向基准给图像的各个局部特征,结合相邻像素点的尺度方向分布特性计算关键像素点的数值与方向,关键像素点的数值与方向计算公式如下:
其中,I(x,y)表示极值点的数值,表示极值点的方向,(x+1,y+1)与(x-1,y-1)表示极值点相邻像素点在尺度空间内的信息表征;
选取关键点周围的像素点区域并划子区域,以每一个子区域的中心像素点为中心在子正方形中做最大内切十六边形,计算来自不同方向的特征向量作为子区域中心点的特征向量,将子区域的特征向量以距离为权重叠加在关键特征点上,生成关键点描述符。
作为本发明所述基于改进特征描述符和KNN搜索的眼底图像特征点匹配方法的一种优选方案,其中:所述判断极值点的具体步骤包括,
通过二维高斯函数与原眼底图像进行卷积构建展示多尺度参数下综合信息的图像尺度空间,使用高斯函数对图像降采样,生成图像金字塔模型;
截取每张图像高斯金字塔前三层,将三层高斯金字塔依次相减得到两层高斯差分金字塔;
将差分金字塔图像的像素点数值与周围相邻尺度范围内的像素点数值进行比较,判断得到极值点。
作为本发明所述基于改进特征描述符和KNN搜索的眼底图像特征点匹配方法的一种优选方案,其中:所述消除误差匹配包括以下步骤,
根据KNN算法对输入关键点描述符进行最近邻特征点与次近邻特征点的搜索;
计算特征点最近邻距离与次近邻距离的距离比值,根据不同图像计算得到正比阈值,相关计算公式如下:
其中,R表示正比阈值,N表示特征点总数,wk表示第k个特征点的权重,d(xk,yk)表示第k个特征点与其最近邻点间的欧式距离,α表示归一化系数,wj表示对所有特征点权重进行归一化的权重系数;
当最近邻距离与次近邻距离的距离比值小于正比阈值时,则表示匹配点对符合特征点需求,接受匹配点对;
当最近邻距离与次近邻距离的距离比值大于正比阈值时,则表示匹配点对不符合特征点需求,将匹配点对剔除。
第二方面,本发明提供了基于改进特征描述符和KNN搜索的眼底图像特征点匹配系统,其包括:所述系统包括图像预处理模块、特征点检测模块、特征描述符生成模块以及特征匹配模块;
图像预处理模块用于获取眼底图像,转化为灰度图像,进行直方图均衡化和双边滤波增强,得到眼底表征图像;
特征点检测模块用于获得改进的尺度空间和高斯差分金字塔技术,比较像素点值判断关键像素点,计算点的位置和方向信息;
特征描述符生成模块用于围绕关键点划分子区域,计算每个子区域的特征向量,并结合距离权重叠加生成关键点的特征描述符;
特征匹配模块用于采用KNN近邻搜索算法找到每个关键点的最近邻匹配点和次近邻匹配点,计算距离比值消除误匹配。
第三方面,本发明实施例提供了一种计算机设备,包括存储器和处理器,所述存储器存储有计算机程序,其中:所述计算机程序被处理器执行时实现如本发明第一方面所述的基于改进特征描述符和KNN搜索的眼底图像特征点匹配方法的任一步骤。
第四方面,本发明实施例提供了一种计算机可读存储介质,其上存储有计算机程序,其中:所述计算机程序被处理器执行时实现如本发明第一方面所述的基于改进特征描述符和KNN搜索的眼底图像特征点匹配方法的任一步骤。
本发明有益效果为实现了对眼底图像的高效处理和特征提取,为后续匹配提供了可靠的基础,有助于提高匹配的准确性和鲁棒性。通过优化图像预处理,提高了图像的质量和表征能力,为后续步骤提供更好的输入,有助于增强眼底表征图像的特征,提高匹配的可靠性。通过生成均匀分布的眼底图像直方图,优化了图像的灰度分布,有助于提高图像的对比度和信息表达能力,为特征提取提供更有区分度的图像表示。通过双边滤波增强步骤,点距权重与点值权重的计算实现了保持图像边缘细节的同时平滑图像的双边滤波效果,进一步提升了图像的质量。通过描述符生成步骤,子区域特征向量的距离权重叠加方法结合了局部和全局信息,使描述符对场景变化更加鲁棒。总体来说,本发明中的多个步骤相互配合,使图像关键点匹配效果显著提升,通过距离比值判断机制消除了误匹配,保证了匹配结果的可靠性。
为了更清楚地说明本发明实施例的技术方案,下面将对实施例描述中所需要使用的附图作简单地介绍,显而易见地,下面描述中的附图仅仅是本发明的一些实施例,对于本领域普通技术人员来讲,在不付出创造性劳动的前提下,还可以根据这些附图获得其它的附图。其中:
图1为基于改进特征描述符和KNN搜索的眼底图像特征点匹配方法的场景图。
图2为基于改进特征描述符和KNN搜索的眼底图像特征点匹配方法的结构图。
为使本发明的上述目的、特征和优点能够更加明显易懂,下面结合说明书附图对本发明的具体实施方式做详细的说明。
在下面的描述中阐述了很多具体细节以便于充分理解本发明,但是本发明还可以采用其他不同于在此描述的其它方式来实施,本领域技术人员可以在不违背本发明内涵的情况下做类似推广,因此本发明不受下面公开的具体实施例的限制。
其次,此处所称的“一个实施例”或“实施例”是指可包含于本发明至少一个实现方式中的特定特征、结构或特性。在本说明书中不同地方出现的“在一个实施例中”并非均指同一个实施例,也不是单独的或选择性的与其他实施例互相排斥的实施例。
实施例1
参照图1和图2,为本发明第一个实施例,该实施例提供了基于改进特征描述符和KNN搜索的眼底图像特征点匹配方法,
输入需要拼接的眼底图像集,对图片集进行图像预处理,获得眼底表征图像。
进一步的,图像预处理包括以下步骤,
获取眼底图像数据集;
输入眼底图像并转化为灰度图像,求出眼底图像灰度直方图,计算灰度值变换表,根据给定眼底图像灰度直方图生成均匀分布的眼底图像直方图;
根据灰度值变换表计算均匀分布的眼底图像直方图中各像素中心周围的像素点灰度值与空域值,对像素点灰度值与空域值进行加权平均,进行双边滤波过滤增强,得到眼底表征图像。
进一步的,生成均匀分布的眼底图像直方图包括以下步骤,
将眼底图像转化为灰度图像,并扫描原始灰度图像的每一个像素点,得出眼底图像灰度直方图;
计算眼底图像灰度直方图的累积分布函数,计算公式如下:
其中,n表示一幅图像的像素点数目,l表示灰度级总数,nk表示灰度级为rk的像素点数目;
根据直方图灰度值均匀原理和累积分布函数计算映射函数,计算公式如下:
根据映射函数计算出每个像素点经过映射后的数值,生成均匀分布的眼底图像直方图。
更近一步的,双边滤波过滤增强的具体步骤包括,
输入经过直方图均匀后的眼底图像,针对每张图像像素点计算双边滤波器中的点距权重与点值权重,计算公式如下:
其中,Wd表示点距权重,Wi表示点值权重,||p-q||是点p与点q之间的欧式距离,|Ip-Iq|是两点的像素差值,表示像素方差,表示距离方差;
根据点距权重与点值权重计算图像上每个点的融合权重,计算公式如下:
其中,Wp表示融合之后的权重,A表示计算的图像像素点集;
通过每个像素点周边像素的加权平均值来计算过滤后的像素强度,计算公式如下:
其中,I'p表示过滤后的像素强度。
构建改进的图像尺度空间金字塔,并运用高斯差分金字塔技术比较眼底表征图像像素值初步判定极值点,定位关键像素点的位置与方向,并采用改进特征向量生成特征描述符。
进一步的,生成特征描述符的具体步骤包括,
根据眼底表征图像构建高斯金字塔,取每组图像金字塔的前三层计算差分图像,生成高斯差分金字塔,通过高斯差分金字塔与原图像卷积得到改进尺度空间图像,比较金字塔中区域内像素点的像素值判断极值点;
根据极值点的尺度不变性,分配方向基准给图像的各个局部特征,结合相邻像素点的尺度方向分布特性计算关键像素点的数值与方向,关键像素点的数值与方向计算公式如下:
其中,I(x,y)表示极值点的数值,表示极值点的方向,(x+1,y+1)与(x-1,y-1)表示极值点相邻像素点在尺度空间内的信息表征;
选取关键点周围的像素点区域并划子区域,以每一个子区域的中心像素点为中心在子正方形中做最大内切十六边形,计算来自不同方向的特征向量作为子区域中心点的特征向量,将子区域的特征向量以距离为权重叠加在关键特征点上,生成关键点描述符。
具体的,划分关键像素点周围的子区域为4个子区域,定义每个子区域包括16个方向特征,结合生成64维向量的关键点描述符,来表征眼底图像关键像素点的特征信息用以特征匹配。
更进一步的,判断极值点的具体步骤包括,
通过二维高斯函数与原眼底图像进行卷积构建展示多尺度参数下综合信息的图像尺度空间,使用高斯函数对图像降采样,生成图像金字塔模型;
截取每张图像高斯金字塔前三层,将三层高斯金字塔依次相减得到两层高斯差分金字塔;
将差分金字塔图像的像素点数值与周围相邻尺度范围内的像素点数值进行比较,判断得到极值点。
通过KNN算法搜索每组眼底图像中关键点的最近邻与次近邻匹配点,消除误差匹配,保留优秀匹配对。
进一步的,消除误差匹配包括以下步骤,
根据KNN算法对输入关键点描述符进行最近邻特征点与次近邻特征点的搜索;
计算特征点最近邻距离与次近邻距离的距离比值,根据不同图像计算得到正比阈值,相关计算公式如下:
其中,R表示正比阈值,N表示特征点总数,wk表示第k个特征点的权重,d(xk,yk)表示第k个特征点与其最近邻点间的欧式距离,α表示归一化系数,wj表示对所有特征点权重进行归一化的权重系数;
当最近邻距离与次近邻距离的距离比值小于正比阈值时,则表示匹配点对符合特征点需求,接受匹配点对;
当最近邻距离与次近邻距离的距离比值大于正比阈值时,则表示匹配点对不符合特征点需求,将匹配点对剔除。
具体的,KNN算法中的k值取2。
本实施例还提供一种计算机设备,适用于基于改进特征描述符和KNN搜索的眼底图像特征点匹配方法的情况,包括:存储器和处理器;存储器用于存储计算机可执行指令,处理器用于执行计算机可执行指令,实现如上述实施例提出的基于改进特征描述符和KNN搜索的眼底图像特征点匹配方法。
该计算机设备可以是终端,该计算机设备包括通过系统总线连接的处理器、存储器、通信接口、显示屏和输入装置。其中,该计算机设备的处理器用于提供计算和控制能力。该计算机设备的存储器包括非易失性存储介质、内存储器。该非易失性存储介质存储有操作系统和计算机程序。该内存储器为非易失性存储介质中的操作系统和计算机程序的运行提供环境。该计算机设备的通信接口用于与外部的终端进行有线或无线方式的通信,无线方式可通过WIFI、运营商网络、NFC(近场通信)或其他技术实现。该计算机设备的显示屏可以是液晶显示屏或者电子墨水显示屏,该计算机设备的输入装置可以是显示屏上覆盖的触摸层,也可以是计算机设备外壳上设置的按键、轨迹球或触控板,还可以是外接的键盘、触控板或鼠标等。
本实施例还提供一种存储介质,其上存储有计算机程序,该程序被处理器执行时实现如上述实施例提出的实现基于改进特征描述符和KNN搜索的眼底图像特征点匹配方法;存储介质可以由任何类型的易失性或非易失性存储设备或者它们的组合实现,如静态随机存取存储器(Static Random Access Memory,简称SRAM),电可擦除可编程只读存储器(Electrically Erasable Programmable Read-Only Memory,简称EEPROM),可擦除可编程只读存储器(Erasable Programmable Read Only Memory,简称EPROM),可编程只读存储器(Programmable Red-Only Memory,简称PROM),只读存储器(Read-Only Memory,简称ROM),磁存储器,快闪存储器,磁盘或光盘。
综上,本发明实现了对眼底图像的高效处理和特征提取,为后续匹配提供了可靠的基础,有助于提高匹配的准确性和鲁棒性。通过优化图像预处理,提高了图像的质量和表征能力,为后续步骤提供更好的输入,有助于增强眼底表征图像的特征,提高匹配的可靠性。通过生成均匀分布的眼底图像直方图,优化了图像的灰度分布,有助于提高图像的对比度和信息表达能力,为特征提取提供更有区分度的图像表示。通过双边滤波增强步骤,点距权重与点值权重的计算实现了保持图像边缘细节的同时平滑图像的双边滤波效果,进一步提升了图像的质量。通过描述符生成步骤,子区域特征向量的距离权重叠加方法结合了局部和全局信息,使描述符对场景变化更加鲁棒。总体来说,本发明中的多个步骤相互配合,使图像关键点匹配效果显著提升,通过距离比值判断机制消除了误匹配,保证了匹配结果的可靠性。
实施例2
参照图1-图2,为本发明第二个实施例,该实施例提供了基于改进特征描述符和KNN搜索的眼底图像特征点匹配方法,为了验证本发明的有益效果,通过经济效益计算和仿真实验进行科学论证。
试验准备
数据集准备:选择两组眼底图像数据集,一组用于应用本发明方法,另一组用于应用传统SIFT算法。
软件和硬件环境:准备具有高性能计算能力的计算机,并安装适当的图像处理软件和算法库。
实施详细过程
图像预处理:对两组眼底图像数据集进行灰度转换、直方图均衡化和双边滤波处理。
特征点检测与描述:在本发明方法中,构建改进的尺度空间金字塔,使用高斯差分金字塔技术确定极值点,然后生成改进的特征描述符。在传统SIFT方法中,采用标准的特征检测和描述算法。
特征点匹配:使用KNN算法进行特征点匹配,本发明方法中加入了距离比值判断机制来减少误匹配,而传统SIFT算法则直接进行匹配。
根据表1对比数据表进行数据判断。
表1对比数据表
对比本发明方法与传统SIFT算法,从表格数据可以看出,本发明方法在特征点匹配的精度上平均提高了8.7%,在处理速度上则平均快了0.7秒。这一显著提升主要归功于以下几点创新:
改进的尺度空间金字塔:更有效地捕捉图像特征,增强了特征点的检测精度。
优化的特征描述符:更准确地描述特征点,增强了匹配过程中的准确性。
KNN算法结合距离比值判断机制:有效减少误匹配,提高了整体的匹配准确率。
本发明的实施例展示了其在眼底图像特征点匹配中的优势,特别是在匹配准确性和处理速度方面。这些优势对于眼底图像的分析具有重要意义,提高了算法在实际应用中的可靠性和效率。
应说明的是,以上实施例仅用以说明本发明的技术方案而非限制,尽管参照较佳实施例对本发明进行了详细说明,本领域的普通技术人员应当理解,可以对本发明的技术方案进行修改或者等同替换,而不脱离本发明技术方案的精神和范围,其均应涵盖在本发明的权利要求范围当中。
Claims (8)
- 基于改进特征描述符和KNN搜索的眼底图像特征点匹配方法,其特征在于:包括,输入需要拼接的眼底图像集并进行图像预处理,获得眼底表征图像;构建改进的图像尺度空间金字塔,并运用高斯差分金字塔技术比较眼底表征图像像素值初步判定极值点,定位关键像素点的位置与方向,采用改进特征向量生成特征描述符;所述生成特征描述符的具体步骤包括,根据眼底表征图像构建高斯金字塔,取每组图像金字塔的前三层计算差分图像,生成高斯差分金字塔,通过高斯差分金字塔与原图像卷积得到改进尺度空间图像,比较金字塔中区域内像素点的像素值判断极值点;根据极值点的尺度不变性,分配方向基准给图像的各个局部特征,结合相邻像素点的尺度方向分布特性计算关键像素点的数值与方向,关键像素点的数值与方向计算公式如下:
其中,I(x,y)表示极值点的数值,表示极值点的方向,(x+1,y+1)与(x-1,y-1)表示极值点相邻像素点在尺度空间内的信息表征;选取关键点周围的像素点区域并划子区域,以每一个子区域的中心像素点为中心在子正方形中做最大内切十六边形,计算来自不同方向的特征向量作为子区域中心点的特征向量,将子区域的特征向量以距离为权重叠加在关键特征点上,生成关键点描述符;通过KNN算法搜索每组眼底图像中关键点的最近邻与次近邻匹配点,消除误差匹配,保留优秀匹配对;所述消除误差匹配包括以下步骤,根据KNN算法对输入关键点描述符进行最近邻特征点与次近邻特征点的搜索;计算特征点最近邻距离与次近邻距离的距离比值,根据不同图像计算得到正比阈值,相关计算公式如下:
其中,R表示正比阈值,N表示特征点总数,wk表示第k个特征点的权重,d(xk,yk)表示第k个特征点与其最近邻点间的欧式距离,α表示归一化系数,wj表示对所有特征点权重进行归一化的权重系数;当最近邻距离与次近邻距离的距离比值小于正比阈值时,则表示匹配点对符合特征点需求,接受匹配点对;当最近邻距离与次近邻距离的距离比值大于正比阈值时,则表示匹配点对不符合特征点需求,将匹配点对剔除。 - 如权利要求1所述的基于改进特征描述符和KNN搜索的眼底图像特征点匹配方法,其特征在于:所述图像预处理包括以下步骤,获取眼底图像数据集;输入眼底图像并转化为灰度图像,求出眼底图像灰度直方图,计算灰度值变换表,根据给定眼底图像灰度直方图生成均匀分布的眼底图像直方图;根据灰度值变换表计算均匀分布的眼底图像直方图中各像素中心周围的像素点灰度值与空域值,对像素点灰度值与空域值进行加权平均,进行双边滤波过滤增强,得到眼底表征图像。
- 如权利要求2所述的基于改进特征描述符和KNN搜索的眼底图像特征点匹配方法,其特征在于:所述生成均匀分布的眼底图像直方图包括以下步骤,将眼底图像转化为灰度图像,并扫描原始灰度图像的每一个像素点,得出眼底图像灰度直方图;计算眼底图像灰度直方图的累积分布函数,计算公式如下:
其中,n表示一幅图像的像素点数目,l表示灰度级总数,nk表示灰度级为rk的像素点数目;根据直方图灰度值均匀原理和累积分布函数计算映射函数,计算公式如下:
根据映射函数计算出每个像素点经过映射后的数值,生成均匀分布的眼底图像直方图。 - 如权利要求2所述的基于改进特征描述符和KNN搜索的眼底图像特征点匹配方法,其特征在于:所述双边滤波过滤增强的具体步骤包括,输入经过直方图均匀后的眼底图像,针对每张图像像素点计算双边滤波器中的点距权重与点值权重,计算公式如下:
其中,Wd表示点距权重,Wi表示点值权重,||p-q||是点p与点q之间的欧式距离,|Ip-Iq|是两点的像素差值,表示像素方差,表示距离方差;根据点距权重与点值权重计算图像上每个点的融合权重,计算公式如下:
其中,Wp表示融合之后的权重,A表示计算的图像像素点集;通过每个像素点周边像素的加权平均值来计算过滤后的像素强度,计算公式如下:
其中,I'p表示过滤后的像素强度。 - 如权利要求1所述的基于改进特征描述符和KNN搜索的眼底图像特征点匹配方法,其特征在于:所述判断极值点的具体步骤包括,通过二维高斯函数与原眼底图像进行卷积构建展示多尺度参数下综合信息的图像尺度空间,使用高斯函数对图像降采样,生成图像金字塔模型;截取每张图像高斯金字塔前三层,将三层高斯金字塔依次相减得到两层高斯差分金字塔;将差分金字塔图像的像素点数值与周围相邻尺度范围内的像素点数值进行比较,判断得到极值点。
- 基于改进特征描述符和KNN搜索的眼底图像特征点匹配系统,基于权利要求1~5任一所述的基于改进特征描述符和KNN搜索的眼底图像特征点匹配方法,其特征在于:所述系统包括图像预处理模块、特征点检测模块、特征描述符生成模块以及特征匹配模块;图像预处理模块用于获取眼底图像,转化为灰度图像,进行直方图均衡化和双边滤波增强,得到眼底表征图像;特征点检测模块用于获得改进的尺度空间和高斯差分金字塔技术,比较像素点值判断关键像素点,计算点的位置和方向信息;特征描述符生成模块用于围绕关键点划分子区域,计算每个子区域的特征向量,并结合距离权重叠加生成关键点的特征描述符;特征匹配模块用于采用KNN近邻搜索算法找到每个关键点的最近邻匹配点和次近邻匹配点,计算距离比值消除误匹配。
- 一种计算机设备,包括存储器和处理器,所述存储器存储有计算机程序,其特征在于:所述处理器执行所述计算机程序时实现权利要求1~5任一所述的基于改进特征描述符和KNN搜索的眼底图像特征点匹配方法的步骤。
- 一种计算机可读存储介质,其上存储有计算机程序,其特征在于:所述计算机程序被处理器执行时实现权利要求1~5任一所述的基于改进特征描述符和KNN搜索的眼底图像特征点匹配方法的步骤。
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410373152.X | 2024-03-29 | ||
| CN202410373152.XA CN117975546B (zh) | 2024-03-29 | 2024-03-29 | 基于改进特征描述符和knn搜索的眼底图像特征点匹配方法 |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| WO2025200534A1 true WO2025200534A1 (zh) | 2025-10-02 |
Family
ID=90858099
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| PCT/CN2024/135993 Pending WO2025200534A1 (zh) | 2024-03-29 | 2024-12-02 | 基于改进特征描述符和knn搜索的眼底图像特征点匹配方法 |
Country Status (2)
| Country | Link |
|---|---|
| CN (1) | CN117975546B (zh) |
| WO (1) | WO2025200534A1 (zh) |
Families Citing this family (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117975546B (zh) * | 2024-03-29 | 2024-06-14 | 南京邮电大学 | 基于改进特征描述符和knn搜索的眼底图像特征点匹配方法 |
| CN119671984B (zh) * | 2024-11-29 | 2025-12-26 | 肇庆佳旺纸品包装有限公司 | 一种运用图像处理检测包装盒瑕疵的方法及系统 |
| CN119723123B (zh) * | 2025-02-27 | 2025-06-06 | 苏州元脑智能科技有限公司 | 图片相似度检测方法、装置、设备及存储介质 |
| CN120431137B (zh) * | 2025-04-30 | 2025-10-28 | 中国人民解放军国防科技大学 | 基于机器学习的关联成像信息提取技术及并行采样系统 |
| CN121280229B (zh) * | 2025-12-09 | 2026-03-24 | 苏州比格威医疗科技有限公司 | 一种多视角桶形畸变眼底图像的同步拼接方法、设备及存储介质 |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106651827A (zh) * | 2016-09-09 | 2017-05-10 | 浙江大学 | 一种基于sift特征的眼底图像配准方法 |
| US20190294918A1 (en) * | 2018-03-20 | 2019-09-26 | Nant Holdings Ip, Llc | Volumetric descriptors |
| CN117975546A (zh) * | 2024-03-29 | 2024-05-03 | 南京邮电大学 | 基于改进特征描述符和knn搜索的眼底图像特征点匹配方法 |
Family Cites Families (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP5820154B2 (ja) * | 2010-07-05 | 2015-11-24 | キヤノン株式会社 | 眼科装置、眼科システム及び記憶媒体 |
| CN103729654A (zh) * | 2014-01-22 | 2014-04-16 | 青岛新比特电子科技有限公司 | 基于改进sift算法的图像匹配检索系统 |
| CN104851094A (zh) * | 2015-05-14 | 2015-08-19 | 西安电子科技大学 | 一种基于rgb-d的slam算法的改进方法 |
| CN108022228A (zh) * | 2016-10-31 | 2018-05-11 | 天津工业大学 | 基于SIFT变换和Otsu匹配的彩色眼底图像拼接方法 |
| EP3939006B1 (en) * | 2019-03-15 | 2025-10-29 | Ikerian AG | Feature point detection |
| CN114565781A (zh) * | 2022-02-25 | 2022-05-31 | 中国人民解放军战略支援部队信息工程大学 | 一种基于旋转不变性的影像匹配方法 |
| CN115393239A (zh) * | 2022-08-30 | 2022-11-25 | 白建豪 | 一种多模态眼底图像配准和融合方法及系统 |
| CN115439327A (zh) * | 2022-09-22 | 2022-12-06 | 桂林理工大学 | 一种改进surf无人机影像快速拼接方法 |
| CN115546883B (zh) * | 2022-11-07 | 2023-02-28 | 浙江省人民医院 | 眼底图像处理系统 |
| CN115878841B (zh) * | 2023-03-03 | 2023-05-16 | 南京邮电大学 | 一种基于改进秃鹰搜索算法的短视频推荐方法及系统 |
-
2024
- 2024-03-29 CN CN202410373152.XA patent/CN117975546B/zh active Active
- 2024-12-02 WO PCT/CN2024/135993 patent/WO2025200534A1/zh active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106651827A (zh) * | 2016-09-09 | 2017-05-10 | 浙江大学 | 一种基于sift特征的眼底图像配准方法 |
| US20190294918A1 (en) * | 2018-03-20 | 2019-09-26 | Nant Holdings Ip, Llc | Volumetric descriptors |
| CN117975546A (zh) * | 2024-03-29 | 2024-05-03 | 南京邮电大学 | 基于改进特征描述符和knn搜索的眼底图像特征点匹配方法 |
Non-Patent Citations (1)
| Title |
|---|
| ZHANG, XINPENG: "Research on Detection Methods for Microaneurysms in Color Fundus Images", CHINESE DOCTORAL DISSERTATIONS FULL-TEXT DATABASE, MEDICINE AND HEALTH SCIENCES, 3 December 2017 (2017-12-03), XP093359389 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN117975546A (zh) | 2024-05-03 |
| CN117975546B (zh) | 2024-06-14 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113826143B (zh) | 特征点检测 | |
| CN117975546B (zh) | 基于改进特征描述符和knn搜索的眼底图像特征点匹配方法 | |
| WO2021196389A1 (zh) | 面部动作单元识别方法、装置、电子设备及存储介质 | |
| CN112651380B (zh) | 人脸识别方法、人脸识别装置、终端设备及存储介质 | |
| CN110334762A (zh) | 一种基于四叉树结合orb和sift的特征匹配方法 | |
| CN111767960A (zh) | 一种应用于图像三维重建的图像匹配方法及系统 | |
| CN110598715B (zh) | 图像识别方法、装置、计算机设备及可读存储介质 | |
| CN113378620B (zh) | 监控视频噪声环境下跨摄像头行人重识别方法 | |
| Zhang et al. | Unsupervised intrinsic image decomposition using internal self-similarity cues | |
| CN111310688A (zh) | 一种基于多角度成像的手指静脉识别方法 | |
| CN115471682A (zh) | 一种基于SIFT融合ResNet50的图像匹配方法 | |
| CN108550165A (zh) | 一种基于局部不变特征的图像匹配方法 | |
| CN111507288A (zh) | 图像检测方法、装置、计算机设备和存储介质 | |
| Feng et al. | Nonlinear deep subspace network for micro-expression recognition | |
| CN118350991A (zh) | 内窥镜图像拼接方法、内窥镜图像拼接装置、存储介质和计算机设备 | |
| CN113723465B (zh) | 一种改进的特征提取方法以及基于该方法的图像拼接方法 | |
| Azaza et al. | Context proposals for saliency detection | |
| CN109978928B (zh) | 一种基于加权投票的双目视觉立体匹配方法及其系统 | |
| CN114419103A (zh) | 一种骨架检测跟踪方法、装置及电子设备 | |
| Huang et al. | Photo realistic image completion via dense correspondence | |
| Pan et al. | Learnable three-dimensional Gabor convolutional network with global affinity attention for hyperspectral image classification | |
| Wu et al. | Few-Shot Medical Image Segmentation with Large Kernel Attention | |
| CN108765384A (zh) | 一种联合流形排序和改进凸包的显著性检测方法 | |
| Ran et al. | Intrinsic color correction for stereo matching | |
| Peng et al. | An evaluation and study of detail contrast preservation and color consistency in decolorization |
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: 24932925 Country of ref document: EP Kind code of ref document: A1 |