技术领域technical field
本发明涉及一种医学图像处理领域，尤其是涉及一种糖尿病视网膜病变眼底照相标准图像生成方法，用于人工智能医疗诊断。The invention relates to the field of medical image processing, in particular to a standard image generation method for fundus photography of diabetic retinopathy, which is used for artificial intelligence medical diagnosis.
背景技术Background technique
糖尿病视网膜病变(Diabetic Retinopathy，DR)是常见致盲性眼病。中国是全球2型糖尿病患者最多的国家，DR的患病率、致盲率也逐年升高，是目前工作年龄人群第一位的致盲性疾病。目前我国DR在糖尿病罹患人群中的患病率为24.7％～37.5％。根据国际糖尿病联盟统计结果显示，截至2015年，我国糖尿病患者约1.1亿人，按此推算我国DR患者约2700万人，随着糖尿病患者的大幅增多，未来的DR患者数量会成倍增长，DR的防治成为越来越重要的社会问题。而据卫计委统计，我国目前有3.2万名眼科医生，其中，从事眼底医疗服务和研究的医生大约800-1000人，相对于1亿多的糖尿病患者来说，眼科医生严重不足，导致我国糖网病筛查率不足10％，要改变这个现状，可通过人工智能诊断清晰标注的DR眼底照片来解决。Diabetic retinopathy (Diabetic Retinopathy, DR) is a common blinding eye disease. China is the country with the largest number of type 2 diabetes patients in the world, and the prevalence and blindness rate of DR are also increasing year by year. It is currently the number one blinding disease among working-age people. Currently, the prevalence of DR among people with diabetes in my country is 24.7%-37.5%. According to the statistics of the International Diabetes Federation, as of 2015, there were about 110 million people with diabetes in my country. Based on this, it is estimated that there are about 27 million people with DR in my country. The prevention and treatment of the disease has become an increasingly important social issue. According to statistics from the Health and Family Planning Commission, there are currently 32,000 ophthalmologists in my country, of which about 800-1,000 are engaged in fundus medical services and research. The screening rate of diabetic retinopathy is less than 10%. To change this situation, it can be solved by artificial intelligence diagnosis and clearly marked DR fundus photos.
目前，87％的糖尿病患者就诊于县级及以下医疗机构。已研发的DR智能辅助诊断模型训练所使用的DR眼底照相图像均来自于具有良好眼底照相设备和良好DR采集水平医师的大型三级甲等医院，成像质量较高。但是承担大量DR筛查任务的基层眼科，专业设备和专业眼科医师的匮乏，使获取的DR眼底照相图像清晰度和摄影角度等都很难达到DR智能辅助诊断输入要求的理想状态。如何由不标准的DR眼底图像生成标准的DR眼底图像，是需要解决的一个问题。At present, 87% of diabetic patients seek medical treatment in medical institutions at or below the county level. The DR fundus photography images used in the developed DR intelligent auxiliary diagnosis model training are all from large tertiary first-class hospitals with good fundus photography equipment and doctors with good DR acquisition level, and the imaging quality is high. However, the lack of professional equipment and professional ophthalmologists in primary ophthalmology departments that undertake a large number of DR screening tasks makes it difficult for the acquired DR fundus photographic images and camera angles to meet the ideal state required for DR intelligent auxiliary diagnosis input. How to generate standard DR fundus images from non-standard DR fundus images is a problem that needs to be solved.
中国专利(申请号：CN201410078378)公开一种糖尿病视网膜病变的眼底图像特征提取方法，该方法包括以下步骤：(1)眼底图像RGB通道选择；(2)眼底图像视盘定位；(3)对眼底图像进行硬性渗出特征及棉絮斑特征提取，若发现至少一种特征，即生成提取后的眼底图像，若未发现特征，则进行微动脉瘤特征和视网膜内出血特征提取，再生成提取后的眼底图像。Chinese patent (application number: CN201410078378) discloses a fundus image feature extraction method for diabetic retinopathy, the method includes the following steps: (1) RGB channel selection of fundus image; (2) optic disk positioning of fundus image; (3) fundus image Extract the features of hard exudation and cotton wool spots. If at least one feature is found, the extracted fundus image will be generated. If no feature is found, the feature of microaneurysm and intraretinal hemorrhage will be extracted, and the extracted fundus image will be generated again. .
中国论文：糖尿病视网膜病变图像的血管提取方法(作者：单玲玉，2015年5月)，该方法包括：(1)对彩色视网膜图像进行预处理，得到血管与背景部分对比度较高的绿通道图像；(2)对绿通道图像分别进行病变检测处理及图像增强；(3)基于步骤(2)得到病变检测结果，然后对病变区域进行分割；(4)基于步骤(2)得到的增强图像，通过对已有视网膜血管分割方法进行对比，然后结合血管的结构特性——越往末梢越细，仅使用一种阈值分割方法不能有效的分割出细小血管，提出了一种基于全局阈值与局部阈值的视网膜血管分割方法对其进行处理，得到初步的血管分割结果；(5)基于步骤(3)得到的硬性渗出分割结果，将其从步骤(4)中的血管分割结果中去除，然后通过对病变区域内断裂的血管根据梯度强度和方向相似的原则进行连接，得到精确的血管分割结果。Chinese paper: Blood Vessel Extraction Method for Diabetic Retinopathy Image (Author: Shan Lingyu, May 2015), the method includes: (1) Preprocessing the color retinal image to obtain the green channel with high contrast between the blood vessel and the background (2) Carry out lesion detection processing and image enhancement on the green channel image respectively; (3) Obtain the lesion detection result based on step (2), and then segment the lesion area; (4) Enhanced image based on step (2) , by comparing the existing retinal blood vessel segmentation methods, and then combining the structural characteristics of blood vessels - the thinner the end, only one threshold segmentation method can not effectively segment small blood vessels, a new method based on global threshold and local The threshold retinal vessel segmentation method processes it to obtain the preliminary vessel segmentation result; (5) based on the hard exudate segmentation result obtained in step (3), remove it from the vessel segmentation result in step (4), and then By connecting the broken blood vessels in the lesion area according to the principle of gradient strength and direction similarity, accurate blood vessel segmentation results are obtained.
然而，上述现有技术存在共同缺陷就是没有对初次采集的图像(也就是不标准眼底图像)进行二次加工处理，以生成的标准图像清晰度达到智能辅助诊断系统要求，提高诊断正确率的目的。通过本技术，可以去除不标准图像中的干扰因素，保留糖尿病视网膜病变眼底照相图像的病变特征，以用于人工智能医疗诊断。However, there is a common defect in the above-mentioned existing technologies, that is, there is no secondary processing of the first-time collected images (that is, non-standard fundus images), so that the definition of the generated standard images can meet the requirements of the intelligent auxiliary diagnosis system and improve the accuracy of diagnosis. . Through this technology, the interference factors in the non-standard images can be removed, and the lesion characteristics of the fundus photographic images of diabetic retinopathy can be preserved, so as to be used for artificial intelligence medical diagnosis.
发明内容Contents of the invention
针对基层医院专业设备和专业眼科医师的匮乏，使获取的DR眼底照相图像清晰度和摄影角度等都很难达到DR智能辅助诊断输入要求的理想状态的问题，本发明将利用深度学习中GAN技术，将基层不同眼科设备获取的DR眼底照相图像优化提升成标准图像，去除不标准图像中的干扰因素，保留糖尿病视网膜病变眼底图像的病变特征，达到DR智能辅助诊断系统的输入图像标准要求，从而获得较好的诊断结果。因此本发明公开了一种糖尿病视网膜病变眼底照相标准图像生成方法，以解决现有技术中存在的不足。其技术方案如下：In view of the lack of professional equipment and professional ophthalmologists in grass-roots hospitals, it is difficult to achieve the ideal state of DR intelligent auxiliary diagnosis input requirements for the acquired DR fundus photographic images and camera angles. This invention will use GAN technology in deep learning , optimize and upgrade the DR fundus photographic images obtained by different ophthalmic equipment at the grassroots level into standard images, remove the interference factors in non-standard images, retain the lesion characteristics of diabetic retinopathy fundus images, and meet the input image standard requirements of the DR intelligent auxiliary diagnosis system, thereby obtain better diagnostic results. Therefore, the present invention discloses a standard image generation method of diabetic retinopathy fundus photography to solve the deficiencies in the prior art. Its technical scheme is as follows:
一种糖尿病视网膜病变眼底照相标准图像生成方法，包括如下步骤：A standard image generation method for fundus photography of diabetic retinopathy, comprising the following steps:
步骤(1)、将采集的不标准眼底图像通过生成模型生成新样本图像；Step (1), generating a new sample image through the generated model of the collected non-standard fundus image;
步骤(2)、新样本图像局部特征提取；Step (2), new sample image local feature extraction;
步骤(3)、将不标准图像局部特征与标准图像局部特征在判别模型中对比，一致则输出新样本图像，即生成的标准图像，不一致则调整新样本图像。Step (3), comparing the local features of the non-standard image with the local features of the standard image in the discriminant model, if they are consistent, a new sample image is output, that is, the generated standard image, and if they are inconsistent, the new sample image is adjusted.
有益效果：本发明提出的方法简单有效，生成的标准图像清晰度达到智能辅助诊断系统要求，提高诊断正确率。Beneficial effects: the method proposed by the invention is simple and effective, and the definition of the generated standard image meets the requirements of the intelligent auxiliary diagnosis system, thereby improving the diagnosis accuracy.
附图说明Description of drawings
图1为没有经过本发明处理的图像。Fig. 1 is an image not processed by the present invention.
图2为经过本发明方法处理后的标准图像。Fig. 2 is a standard image processed by the method of the present invention.
图3为本发明糖尿病视网膜病变眼底照相标准图像生成方法流程框图。Fig. 3 is a flow chart of the standard image generation method for diabetic retinopathy fundus photography according to the present invention.
具体实施方式detailed description
一种糖尿病视网膜病变眼底照相标准图像生成方法，包括如下步骤：A standard image generation method for fundus photography of diabetic retinopathy, comprising the following steps:
步骤(1)、将采集的不标准眼底图像通过生成模型生成新样本图像；Step (1), generating a new sample image through the generated model of the collected non-standard fundus image;
步骤(2)、新样本图像局部特征提取；所述局部特征包括局部方向梯度直方图HOG特征，尺度不变特征变换SIFT特征，局部颜色特征；Step (2), new sample image local feature extraction; Described local feature comprises local orientation gradient histogram HOG feature, scale-invariant feature transformation SIFT feature, local color feature;
步骤(3)、将不标准图像局部特征与标准图像局部特征在判别模型中对比，一致则输出新样本图像，即生成的标准图像，不一致则调整新样本图像。Step (3), comparing the local features of the non-standard image with the local features of the standard image in the discriminant model, if they are consistent, a new sample image is output, that is, the generated standard image, and if they are inconsistent, the new sample image is adjusted.
实施例1Example 1
步骤(1)进一步包括：采用的生成模型为深度学习中生成对抗网络GAN模型，GAN由Ian Goodfellow在2014年提出，其主要思想是，训练一个生成器(Generator，简称G)，从随机噪声或者潜在变量中生成逼真的样本，同时训练一个判别器(Discriminator，简称D)来判别真实数据和生成数据，两者同时训练，直到达到一个纳什均衡——生成器生成的数据与真实样本无差别，判别器也无法正确区分生成数据和真实数据。该模型可将不标准的糖尿病视网膜病变图像生成标准图像；对于GAN模型，其优化问题是一个极小-极大化问题，其目标函数如公式(1)所示；其中x采样于真实数据分布pdata(x)，z采样于先验分布pz(z)(例如高斯噪声分布)，E(·)表示计算期望值。当固定生成网络G的时候，对于判别网络D的优化，可以这样理解：输入来自于真实数据，D优化网络结构使自己输出1，输入来自于生成数据，D优化网络结构使自己输出0；当固定判别网络D的时候，G优化自己的网络使自己输出尽可能和真实数据一样的样本，并且使得生成的样本经过D的判别之后，D输出高概率。Step (1) further includes: the generation model adopted is the GAN model of the generation confrontation network in deep learning. GAN was proposed by Ian Goodfellow in 2014. The main idea is to train a generator (Generator, referred to as G) from random noise or Generate realistic samples from potential variables, and train a discriminator (D for short) to distinguish real data and generated data, and train both at the same time until a Nash equilibrium is reached—the data generated by the generator is indistinguishable from real samples. The discriminator also cannot correctly distinguish generated data from real data. This model can generate standard images from non-standard diabetic retinopathy images; for the GAN model, its optimization problem is a minimization-maximization problem, and its objective function is shown in formula (1); where x is sampled from the real data distribution pdata (x), z is sampled from the prior distribution pz (z) (such as Gaussian noise distribution), and E(·) means to calculate the expected value. When the fixed generation network G is fixed, the optimization of the discriminant network D can be understood as follows: the input comes from real data, D optimizes the network structure to output 1, the input comes from generated data, and D optimizes the network structure to output 0; when When the discriminant network D is fixed, G optimizes its own network so that it outputs samples that are as similar to real data as possible, and makes the generated samples pass D's discrimination, and D outputs high probability.
实施例2Example 2
所述局部方向梯度直方图HOG特征的提取方法如下：首先将图像分成小的连通区域，即：细胞单元，然后采集细胞单元中各像素点的梯度的或边缘的方向直方图，最后把这些直方图组合起来构成特征描述器。The extraction method of the local orientation gradient histogram HOG feature is as follows: first divide the image into small connected areas, that is: cell units, then collect the gradient or edge direction histograms of each pixel in the cell units, and finally combine these histograms The graphs are combined to form a feature descriptor.
所述局部方向梯度直方图HOG特征的提取方法进一步说明如下：将一个局部区域图像进行：The extraction method of the HOG feature of the local orientation gradient histogram is further described as follows: a local area image is performed:
1)灰度化(将图像看做一个x，y，z(灰度)的三维图像)；1) Grayscale (consider the image as a three-dimensional image of x, y, z (grayscale));
2)采用标准化Gamma空间和颜色空间校正法对输入图像进行颜色空间的标准化(归一化)；其目的是调节图像的对比度，降低图像局部的阴影和光照变化所造成的影响，同时可以抑制噪音的干扰。2) Use standardized Gamma space and color space correction method to standardize (normalize) the color space of the input image; the purpose is to adjust the contrast of the image, reduce the influence of local shadows and illumination changes in the image, and suppress noise at the same time interference.
为了减少光照因素的影响，首先需要将整个图像进行标准化(归一化)。在图像的纹理强度中，局部的表层曝光贡献的比重较大，所以，这种压缩处理能够有效地降低图像局部的阴影和光照变化。因为颜色信息作用不大，通常先转化为灰度图；Gamma压缩公式：In order to reduce the influence of illumination factors, it is first necessary to standardize (normalize) the entire image. In the texture intensity of the image, the local surface exposure contributes a large proportion, so this compression process can effectively reduce the local shadow and illumination changes of the image. Because the color information has little effect, it is usually converted into a grayscale image first; Gamma compression formula:
I(x，y)＝I(x，y)gammaI(x,y)=I(x,y)gamma
如可取为：gamma＝1/2If it can be taken as: gamma=1/2
3)计算图像每个像素的梯度(包括大小和方向)；主要是为了捕获轮廓信息，同时进一步弱化光照的干扰。计算图像横坐标和纵坐标方向的梯度，并据此计算每个像素位置的梯度方向值；求导操作不仅能够捕获轮廓，人影和一些纹理信息，还能进一步弱化光照的影响。图像中像素点(x，y)的梯度为：3) Calculate the gradient (including size and direction) of each pixel of the image; mainly to capture contour information and further weaken the interference of light. Calculate the gradient of the abscissa and ordinate of the image, and calculate the gradient direction value of each pixel position accordingly; the derivation operation can not only capture the outline, shadow and some texture information, but also further weaken the influence of illumination. The gradient of the pixel point (x, y) in the image is:
Gx(x，y)＝H(x+1，y)-H(x-1，y)Gx (x, y) = H (x+1, y) - H (x-1, y)
Gy(x，y)＝H(x，y+1)-H(x，y-1)Gy (x, y)=H(x, y+1)-H(x, y-1)
式中Gx(x，y)，Gy(x，y)，H(x，y)分别表示输入图像中像素点(x，y)处的水平方向梯度、垂直方向梯度和像素值。像素点(x，y)处的梯度幅值和梯度方向分别为：In the formula, Gx (x, y), Gy (x, y), H (x, y) represent the horizontal gradient, vertical gradient and pixel value at the pixel point (x, y) in the input image, respectively. The gradient magnitude and gradient direction at the pixel point (x, y) are respectively:
其方法如下：首先用[1，0，1]梯度算子对原图像做卷积运算，得到x方向(水平方向，以向右为正方向)的梯度分量gradscalx，然后用[1，0，-1]T梯度算子对原图像做卷积运算，得到y方向(竖直方向，以向上为正方向)的梯度分量gradscaly。然后再用以上公式计算该像素点的梯度大小和方向。The method is as follows: first use [1, 0, 1] gradient operator to perform convolution operation on the original image to obtain the gradient component gradscalx in the x direction (horizontal direction, with the positive direction to the right), and then use [1, 0, -1] The T gradient operator performs a convolution operation on the original image to obtain the gradient component gradscaly in the y direction (vertical direction, with upward as the positive direction). Then use the above formula to calculate the gradient size and direction of the pixel.
4)将图像划分成小单元(例如6*6像素/单元)；将局部区域图像分成若干个“单元格cell”，采用9个bin的直方图来统计这6*6个像素的梯度信息。也就是将cell的梯度方向360度分成9个方向块，而梯度大小就是作为投影的权值的。4) Divide the image into small units (for example, 6*6 pixels/unit); divide the local area image into several "cells", and use the histogram of 9 bins to count the gradient information of these 6*6 pixels. That is, the 360-degree gradient direction of the cell is divided into 9 direction blocks, and the gradient size is used as the weight of the projection.
5)统计每个单元的梯度直方图(不同梯度的个数)，即可形成每个单元的描述符；5) Count the gradient histogram (the number of different gradients) of each unit to form the descriptor of each unit;
把细胞单元组合成大的块(block)，块内归一化梯度直方图：Combine the cell units into large blocks, and normalize the gradient histogram within the block:
把各个单元组合成大的、空间上连通的块(blocks)。这样，一个块内所有单元的特征向量串联起来便得到该块的HOG特征。这些区间是互有重叠的，这就意味着：每一个单元格的特征会以不同的结果多次出现在最后的特征向量中。我们将归一化之后的块描述符(向量)就称之为HOG描述符。Combine individual units into large, spatially connected blocks. In this way, the feature vectors of all units in a block are concatenated to obtain the HOG feature of the block. These intervals overlap each other, which means that the features of each cell will appear in the final feature vector multiple times with different results. We call the normalized block descriptor (vector) the HOG descriptor.
6)将每几个单元组成一个块(例如3*3个单元/块)，一个块内所有单元的特征描述符串联起来便得到该块的HOG特征描述符。6) Every several units form a block (for example, 3*3 units/block), and the feature descriptors of all units in a block are concatenated to obtain the HOG feature descriptor of the block.
7)将局部区域图像内的所有块的HOG特征描述符串联起来就可以得到该局部区域图像的HOG特征描述符了。这个就是最终的可供分类使用的特征向量了。7) The HOG feature descriptor of the local area image can be obtained by concatenating the HOG feature descriptors of all blocks in the local area image. This is the final feature vector available for classification.
实施例3Example 3
步骤2中SIFT特征的提取方法Extraction method of SIFT feature in step 2
SIFT的全称是Scale Invariant Feature Transform，尺度不变特征变换，包括4个主要步骤：The full name of SIFT is Scale Invariant Feature Transform, scale invariant feature transformation, including 4 main steps:
1)尺度空间的极值检测：搜索所有尺度空间上的图像，通过高斯微分函数来识别潜在的1) Extremum detection in scale space: search for images in all scale spaces, and identify potential
对尺度和选择不变的兴趣点。Points of interest that are invariant to scale and selection.
通常可使用DoG(差分高斯，Differenceof Gaussina)来近似计算高斯拉普拉斯。Usually, DoG (Difference of Gaussina, Difference ofGaussina ) can be used to approximate Laplacian of Gaussian.
设k为相邻两个高斯尺度空间的比例因子，则DoG的定义：Let k be the scale factor of two adjacent Gaussian scale spaces, then the definition of DoG:
D(x，y，σ)＝[G(x，y，kσ)-G(x，y，σ)]*I(x，y)D(x,y,σ)=[G(x,y,kσ)-G(x,y,σ)]*I(x,y)
＝L(x，y，kσ)-L(x，y，σ)=L(x,y,kσ)-L(x,y,σ)
其中，G(x，y，σ)是高斯核函数。σ称为尺度空间因子，它是高斯正态分布的标准差，反映了图像被模糊的程度，其值越大图像越模糊，对应的尺度也就越大。L(x，y，σ)代表着图像的高斯尺度空间。Among them, G(x, y, σ) is a Gaussian kernel function. σ is called the scale space factor, which is the standard deviation of the Gaussian normal distribution, reflecting the degree of blurring of the image, the larger the value, the blurrier the image, and the larger the corresponding scale. L(x, y, σ) represents the Gaussian scale space of the image.
2)除去不好特征点：在离散空间找到真正意义上的极值点，需要设法将不满足条件的点剔除掉，通过尺度空间DoG函数进行曲线拟合寻找极值点，这一步的本质是去掉DoG局部曲率非常不对称的点。2) Remove bad feature points: To find the real extreme points in the discrete space, you need to try to remove the points that do not meet the conditions, and use the scale space DoG function to perform curve fitting to find the extreme points. The essence of this step is Points where the local curvature of the DoG is very asymmetric are removed.
要剔除掉的不符合要求的点主要有两种：There are two main types of points that do not meet the requirements to be eliminated:
a.低对比度的特征点a. Feature points with low contrast
b.不稳定的边缘响应点b. Unstable edge response points
相应处理技术如下：The corresponding processing techniques are as follows:
①剔除低对比度的特征点① Eliminate low-contrast feature points
候选特征点x，其偏移量定义为Δx，其对比度为D(x)的绝对值|D(x)|，对D(x)应用泰勒展开式Candidate feature point x, its offset is defined as Δx, its contrast is the absolute value of D(x) |D(x)|, and Taylor expansion is applied to D(x)
由于x是D(x)的极值点，所以对上式求导并令其为0，得到Since x is the extremum point of D(x), take the derivative of the above formula and set it to 0, and get
然后再把求得的Δx代入到D(x)的泰勒展开式中Then substitute the obtained Δx into the Taylor expansion of D(x)
设对比度的阈值为T，若|D(x^)|≥T，则该特征点保留，否则剔除掉。Let the contrast threshold be T, if |D(x^)|≥T, then the feature point will be kept, otherwise it will be removed.
②、剔除不稳定的边缘响应点②. Eliminate unstable edge response points
在边缘梯度的方向上主曲率值比较大，而沿着边缘方向则主曲率值较小。候选特征点的DoG函数D(x)的主曲率与2×2Hessian矩阵H的特征值成正比。In the direction of the edge gradient, the principal curvature value is relatively large, while along the edge direction, the principal curvature value is small. The principal curvature of the DoG function D(x) of the candidate feature points is proportional to the eigenvalues of the 2×2 Hessian matrix H.
其中，Dxx，Dxy，Dyy是候选点邻域对应位置的差分求得的。Among them, Dxx , Dxy , and Dyy are obtained from the difference of the corresponding positions of the candidate point neighborhood.
只需检测Just check
其中，Tr(H)为矩阵H的迹，Det(H)为矩阵H的行列式。Tγ为阈值。如果上式成立，则剔除该特征点，否则保留。Among them, Tr(H) is the trace of matrix H, and Det(H) is the determinant of matrix H. Tγ is the threshold. If the above formula is true, then remove the feature point, otherwise keep it.
3)特征方向赋值：基于图像局部的梯度方向，分配给每个关键点位置一个或多个方向，后续的所有操作都是对于关键点的方向、尺度和位置进行变换，从而提供这些特征的不变性。以特征点的为中心、以3×1.5σ为半径的领域内计算各个像素点的梯度的幅角和幅值，然后使用直方图对梯度的幅角进行统计。直方图的横轴是梯度的方向，纵轴为梯度方向对应梯度幅值的累加值，直方图中最高峰所对应的方向即为特征点的方向。3) Feature direction assignment: Based on the local gradient direction of the image, one or more directions are assigned to each key point position, and all subsequent operations are to transform the direction, scale and position of the key points, thereby providing different features of these features. transsexual. The argument and magnitude of the gradient of each pixel point are calculated in the field centered on the feature point and the radius is 3×1.5σ, and then the histogram is used to count the argument of the gradient. The horizontal axis of the histogram is the direction of the gradient, and the vertical axis is the cumulative value of the gradient amplitude corresponding to the gradient direction. The direction corresponding to the highest peak in the histogram is the direction of the feature point.
计算以特征点为中心、以3×1.5σ为半径的区域图像的幅角和幅值，每个点L(x，y)的梯度的模m(x，y)以及方向θ(x，y)可通过下面公式求得：Calculate the argument and amplitude of the region image centered on the feature point and with a radius of 3×1.5σ, the modulus m(x, y) of the gradient of each point L(x, y) and the direction θ(x, y ) can be obtained by the following formula:
4)特征点描述在每个特征点周围的邻域内，在选定的尺度上测量图像的局部梯度，这些梯度被变换成一种表示，这种表示允许比较大的局部形状的变形和光照变换。4) Feature point description In the neighborhood around each feature point, the local gradients of the image are measured at a selected scale, and these gradients are transformed into a representation that allows relatively large local shape deformation and illumination transformation.
首先将坐标轴旋转为特征点的方向，以特征点为中心的16×16的窗口的像素的梯度幅值和方向，将窗口内的像素分成16块，每块是其像素内8个方向的直方图统计，共可形成128维的特征向量。First, the coordinate axis is rotated to the direction of the feature point, the gradient magnitude and direction of the pixel of the 16×16 window centered on the feature point, and the pixels in the window are divided into 16 blocks, each block is the direction of 8 directions in its pixel Histogram statistics can form a total of 128-dimensional feature vectors.
实施例4Example 4
步骤2中局部颜色特征的提取方法Extraction method of local color features in step 2
1)用不同大小的窗口遍历新样本图像，形成不同尺寸的局部图像；1) Use windows of different sizes to traverse the new sample image to form partial images of different sizes;
2)在不同窗口内根据标准图像病灶颜色特征，采用彩色图像分割方法，提取局部图像病灶颜色；2) According to the color features of standard image lesions in different windows, the color image segmentation method is used to extract the color of local image lesions;
3)计算病灶颜色像素点所占局部图像比例值，即为该局部图像颜色特征。3) Calculate the proportion value of the local image that the pixel color of the lesion occupies, which is the color feature of the local image.
实施例5Example 5
步骤(3)中的判别模型采用SVM(支持向量机)进行判别The discriminant model in step (3) adopts SVM (Support Vector Machine) to discriminate
将生成的新样本图像提取的局部特征与标准图像局部特征比较，如果其特征差距在一定范围内，则认为是真实的，否则认为是假的。如果为假，则需要修改生成的新样本图像，再次提取其特征进行判别，直到判别真实为止。Compare the local features extracted from the generated new sample image with the local features of the standard image, if the feature gap is within a certain range, it is considered real, otherwise it is considered false. If it is false, you need to modify the generated new sample image, extract its features again for discrimination, until the discrimination is true.
在以上的描述中阐述了很多具体细节以便于充分理解本发明。但是以上描述仅是本发明的较佳实施例而已，本发明能够以很多不同于在此描述的其它方式来实施，因此本发明不受上面公开的具体实施的限制。同时任何熟悉本领域技术人员在不脱离本发明技术方案范围情况下，都可利用上述揭示的方法和技术内容对本发明技术方案做出许多可能的变动和修饰，或修改为等同变化的等效实施例。凡是未脱离本发明技术方案的内容，依据本发明的技术实质对以上实施例所做的任何简单修改、等同变化及修饰，均仍属于本发明技术方案保护的范围内。In the foregoing description, numerous specific details are set forth in order to provide a thorough understanding of the present invention. However, the above descriptions are only preferred embodiments of the present invention, and the present invention can be implemented in many other ways different from those described here, so the present invention is not limited by the specific implementations disclosed above. At the same time, any person skilled in the art can use the methods and technical content disclosed above to make many possible changes and modifications to the technical solution of the present invention without departing from the scope of the technical solution of the present invention, or modify it into an equivalent implementation of equivalent changes example. All the content that does not deviate from the technical solution of the present invention, any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present invention still belong to the protection scope of the technical solution of the present invention.
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710706334.4ACN107506770A (en) | 2017-08-17 | 2017-08-17 | Diabetic retinopathy eye-ground photography standard picture generation method |
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710706334.4ACN107506770A (en) | 2017-08-17 | 2017-08-17 | Diabetic retinopathy eye-ground photography standard picture generation method |
| Publication Number | Publication Date |
|---|---|
| CN107506770Atrue CN107506770A (en) | 2017-12-22 |
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710706334.4APendingCN107506770A (en) | 2017-08-17 | 2017-08-17 | Diabetic retinopathy eye-ground photography standard picture generation method |
| Country | Link |
|---|---|
| CN (1) | CN107506770A (en) |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108510503A (en)* | 2018-03-19 | 2018-09-07 | 哈尔滨工程大学 | A kind of multi-texturing characteristic image dividing method based on Nash Equilibrium Theory |
| CN108537801A (en)* | 2018-03-29 | 2018-09-14 | 山东大学 | Based on the retinal angiomatous image partition method for generating confrontation network |
| CN108765379A (en)* | 2018-05-14 | 2018-11-06 | 深圳明眸科技有限公司 | Computational methods, device, Medical Devices and the storage medium of eyeground pathological changes region area |
| CN108765408A (en)* | 2018-05-31 | 2018-11-06 | 杭州同绘科技有限公司 | Build the method in cancer pathology image virtual case library and the multiple dimensioned cancer detection system based on convolutional neural networks |
| CN109166126A (en)* | 2018-08-13 | 2019-01-08 | 苏州比格威医疗科技有限公司 | A method of paint crackle is divided on ICGA image based on condition production confrontation network |
| CN109300138A (en)* | 2018-08-02 | 2019-02-01 | 哈尔滨工程大学 | A Medical Image Segmentation Method Based on Two-step Nash Equilibrium Improved C-V Model |
| CN109409508A (en)* | 2018-11-06 | 2019-03-01 | 成都信息工程大学 | A method of model avalanche is solved based on confrontation Web vector graphic perception loss is generated |
| CN109508647A (en)* | 2018-10-22 | 2019-03-22 | 北京理工大学 | A kind of spectra database extended method based on generation confrontation network |
| CN109934116A (en)* | 2019-02-19 | 2019-06-25 | 华南理工大学 | A Standard Face Generation Method Based on Generative Adversarial Mechanism and Attention Mechanism |
| CN109978796A (en)* | 2019-04-04 | 2019-07-05 | 北京百度网讯科技有限公司 | Optical fundus blood vessel Picture Generation Method, device and storage medium |
| CN110211087A (en)* | 2019-01-28 | 2019-09-06 | 南通大学 | The semi-automatic diabetic eyeground pathological changes mask method that can share |
| CN110222728A (en)* | 2019-05-15 | 2019-09-10 | 图灵深视(南京)科技有限公司 | The training method of article discrimination model, system and article discrimination method, equipment |
| CN110517248A (en)* | 2019-08-27 | 2019-11-29 | 北京百度网讯科技有限公司 | Fundus image processing, training method, device and equipment thereof |
| CN110674828A (en)* | 2018-07-03 | 2020-01-10 | 江威 | Method and device for normalizing fundus images |
| CN110751187A (en)* | 2019-09-26 | 2020-02-04 | 上海联影智能医疗科技有限公司 | Training method of abnormal area image generation network and related product |
| CN111080524A (en)* | 2019-12-19 | 2020-04-28 | 吉林农业大学 | Identification method of plant diseases and insect pests based on deep learning |
| CN111325809A (en)* | 2020-02-07 | 2020-06-23 | 广东工业大学 | Appearance image generation method based on double-impedance network |
| CN111971689A (en)* | 2018-04-13 | 2020-11-20 | 医科达有限公司 | Image synthesis using countermeasure networks |
| JP2021079042A (en)* | 2019-11-22 | 2021-05-27 | キヤノン株式会社 | Image processing device, image processing method, and program |
| CN115985472A (en)* | 2022-12-01 | 2023-04-18 | 珠海全一科技有限公司 | Fundus image labeling method and system based on neural network |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106709958A (en)* | 2016-12-03 | 2017-05-24 | 浙江大学 | Gray scale gradient and color histogram-based image quality evaluation method |
| CN106951867A (en)* | 2017-03-22 | 2017-07-14 | 成都擎天树科技有限公司 | Face recognition method, device, system and equipment based on convolutional neural network |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106709958A (en)* | 2016-12-03 | 2017-05-24 | 浙江大学 | Gray scale gradient and color histogram-based image quality evaluation method |
| CN106951867A (en)* | 2017-03-22 | 2017-07-14 | 成都擎天树科技有限公司 | Face recognition method, device, system and equipment based on convolutional neural network |
| Title |
|---|
| BROOK_ICV: "《SIFT概述》", 《HTTPS://WWW.CNBLOGS.COM/WANGGUCHANGQING/P/4853263.HTML》* |
| IAN J. GOODFELLOW等: "《Generative Adversarial Nets》", 《ARXIV:1406.2661 [CS, STAT]》* |
| RANJIEWEN: "《HOG：用于人体检测的梯度方向直方图 Histograms of Oriented Gradients for Human Detection》", 《HTTPS://WWW.CNBLOGS.COM/RANJIEWEN/ARTICLES/6820541.HTML》* |
| 王坤峰等: "《生成式对抗网络GAN的研究进展与展望》", 《自动化学报》* |
| 郭晓新等: "《视网膜血流造影图像序列的超分辨率重建》", 《吉林大学学报（理学版）》* |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108510503B (en)* | 2018-03-19 | 2021-05-14 | 哈尔滨工程大学 | A Multi-Texture Feature Image Segmentation Method Based on Nash Equilibrium Theory |
| CN108510503A (en)* | 2018-03-19 | 2018-09-07 | 哈尔滨工程大学 | A kind of multi-texturing characteristic image dividing method based on Nash Equilibrium Theory |
| CN108537801A (en)* | 2018-03-29 | 2018-09-14 | 山东大学 | Based on the retinal angiomatous image partition method for generating confrontation network |
| CN111971689A (en)* | 2018-04-13 | 2020-11-20 | 医科达有限公司 | Image synthesis using countermeasure networks |
| CN108765379B (en)* | 2018-05-14 | 2019-11-19 | 深圳明眸科技有限公司 | Calculation method, device, Medical Devices and the storage medium of eyeground pathological changes region area |
| CN108765379A (en)* | 2018-05-14 | 2018-11-06 | 深圳明眸科技有限公司 | Computational methods, device, Medical Devices and the storage medium of eyeground pathological changes region area |
| CN108765408A (en)* | 2018-05-31 | 2018-11-06 | 杭州同绘科技有限公司 | Build the method in cancer pathology image virtual case library and the multiple dimensioned cancer detection system based on convolutional neural networks |
| CN110674828A (en)* | 2018-07-03 | 2020-01-10 | 江威 | Method and device for normalizing fundus images |
| CN109300138A (en)* | 2018-08-02 | 2019-02-01 | 哈尔滨工程大学 | A Medical Image Segmentation Method Based on Two-step Nash Equilibrium Improved C-V Model |
| CN109166126B (en)* | 2018-08-13 | 2022-02-18 | 苏州比格威医疗科技有限公司 | Method for segmenting paint cracks on ICGA image based on condition generation type countermeasure network |
| CN109166126A (en)* | 2018-08-13 | 2019-01-08 | 苏州比格威医疗科技有限公司 | A method of paint crackle is divided on ICGA image based on condition production confrontation network |
| CN109508647A (en)* | 2018-10-22 | 2019-03-22 | 北京理工大学 | A kind of spectra database extended method based on generation confrontation network |
| CN109409508B (en)* | 2018-11-06 | 2022-03-15 | 成都信息工程大学 | Method for solving model collapse based on generation of confrontation network use perception loss |
| CN109409508A (en)* | 2018-11-06 | 2019-03-01 | 成都信息工程大学 | A method of model avalanche is solved based on confrontation Web vector graphic perception loss is generated |
| CN110211087B (en)* | 2019-01-28 | 2023-06-30 | 南通大学 | Shareable semi-automatic diabetic retinopathy labeling method |
| CN110211087A (en)* | 2019-01-28 | 2019-09-06 | 南通大学 | The semi-automatic diabetic eyeground pathological changes mask method that can share |
| CN109934116A (en)* | 2019-02-19 | 2019-06-25 | 华南理工大学 | A Standard Face Generation Method Based on Generative Adversarial Mechanism and Attention Mechanism |
| CN109978796B (en)* | 2019-04-04 | 2021-06-01 | 北京百度网讯科技有限公司 | Fundus blood vessel picture generation method and device and storage medium |
| CN109978796A (en)* | 2019-04-04 | 2019-07-05 | 北京百度网讯科技有限公司 | Optical fundus blood vessel Picture Generation Method, device and storage medium |
| CN110222728B (en)* | 2019-05-15 | 2021-03-12 | 图灵深视(南京)科技有限公司 | Training method and system of article identification model and article identification method and equipment |
| CN110222728A (en)* | 2019-05-15 | 2019-09-10 | 图灵深视(南京)科技有限公司 | The training method of article discrimination model, system and article discrimination method, equipment |
| CN110517248A (en)* | 2019-08-27 | 2019-11-29 | 北京百度网讯科技有限公司 | Fundus image processing, training method, device and equipment thereof |
| CN110751187B (en)* | 2019-09-26 | 2022-08-19 | 武汉中科医疗科技工业技术研究院有限公司 | Training method of abnormal area image generation network and related product |
| CN110751187A (en)* | 2019-09-26 | 2020-02-04 | 上海联影智能医疗科技有限公司 | Training method of abnormal area image generation network and related product |
| JP7254682B2 (en) | 2019-11-22 | 2023-04-10 | キヤノン株式会社 | Image processing device, image processing method, and program |
| JP2021079042A (en)* | 2019-11-22 | 2021-05-27 | キヤノン株式会社 | Image processing device, image processing method, and program |
| CN111080524A (en)* | 2019-12-19 | 2020-04-28 | 吉林农业大学 | Identification method of plant diseases and insect pests based on deep learning |
| CN111325809A (en)* | 2020-02-07 | 2020-06-23 | 广东工业大学 | Appearance image generation method based on double-impedance network |
| CN111325809B (en)* | 2020-02-07 | 2021-03-12 | 广东工业大学 | Appearance image generation method based on double-impedance network |
| CN115985472A (en)* | 2022-12-01 | 2023-04-18 | 珠海全一科技有限公司 | Fundus image labeling method and system based on neural network |
| CN115985472B (en)* | 2022-12-01 | 2023-09-22 | 珠海全一科技有限公司 | Fundus image labeling method and fundus image labeling system based on neural network |
| Publication | Publication Date | Title |
|---|---|---|
| CN107506770A (en) | Diabetic retinopathy eye-ground photography standard picture generation method | |
| CN108615051A (en) | Diabetic retina image classification method based on deep learning and system | |
| CN108615239B (en) | Tongue Image Segmentation Method Based on Thresholding Technique and Grayscale Projection | |
| CN105513077A (en) | System for screening diabetic retinopathy | |
| CN101246544A (en) | Iris positioning method based on boundary point search and SUSAN edge detection | |
| CN101359365A (en) | Iris positioning method based on maximum between-class variance and gray scale information | |
| Gui et al. | Optic disc localization algorithm based on improved corner detection | |
| CN114300099B (en) | Allolymphocyte typing method based on YOLOv5 and microscopic hyperspectral image | |
| CN106023151A (en) | Traditional Chinese medicine tongue manifestation object detection method in open environment | |
| Hao et al. | Anterior chamber angles classification in anterior segment OCT images via multi-scale regions convolutional neural networks | |
| CN110705468A (en) | Eye movement range identification method and system based on image analysis | |
| CN109033936A (en) | A kind of cervical exfoliated cell core image-recognizing method | |
| CN106778499A (en) | A kind of method of quick positioning people's eye iris during iris capturing | |
| CN106780439A (en) | A method for screening fundus images | |
| CN109816665B (en) | A method and device for fast segmentation of optical coherence tomography images | |
| Aruchamy et al. | Automated glaucoma screening in retinal fundus images | |
| CN114066850A (en) | An Image Binarization Method Based on Classification Framework | |
| CN109948706B (en) | Micro-calcification cluster detection method combining deep learning and feature multi-scale fusion | |
| CN108230341A (en) | A kind of eye fundus image blood vessel segmentation method that nomography is scratched based on layering | |
| WO2022041710A1 (en) | Image-based motion detection method | |
| CN107832695A (en) | The optic disk recognition methods based on textural characteristics and device in retinal images | |
| Joykutty et al. | Automatic tuberculosis detection using adaptive thresholding in chest radiographs | |
| CN108629780B (en) | Tongue image segmentation method based on color decomposition and threshold technology | |
| Rajan et al. | An analytical method for the detection of exudates in retinal images using invertible orientation scores | |
| CN103366376B (en) | A kind of image characteristic extracting method based on neighborhood dimensional variation |
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | Application publication date:20171222 | |
| RJ01 | Rejection of invention patent application after publication |