Referenceless image evaluation method for capsule endoscope, electronic device, and medium

1 . A referenceless image evaluation method for a capsule endoscope, comprising:

inputting original images into a preset image quality evaluation model and a preset image content evaluation model, so as to acquire image quality evaluation scores and image content evaluation scores corresponding to the original images; and

determining, according to weights of the image content evaluation scores and weights of the image quality evaluation scores, a comprehensive score for each image currently being evaluated;

wherein a method for constructing the image quality evaluation model comprises:

analyzing each of the original images to extract a corresponding image quality evaluation characteristic value, wherein the image quality evaluation characteristic value comprises:

at least one of a proportion fb1 of a first overexposed pixel,

a proportion fb2 of a first dark pixel,

a proportion fb3 of a high-frequency coefficient, and

a characteristic value f bri obtained by a BRISQUE (Blind/Referenceless Image Spatial Quality Evaluator) algorithm; and

dividing each of the original images into a first training set and a first testing set according to a predetermined ratio, training data of the first training set by using a support vector machine, and

verifying by using data of the first testing set to obtain the image quality evaluation model,

wherein, both the data of the first training set and the data of the first testing set comprise image quality calculation scores and image quality evaluation characteristic values corresponding to the original images;

wherein before analyzing each of the original image to extract the corresponding image quality evaluation characteristic value, the method further comprises:

clipping the original image at a preset size [W, H] with the center of the original image as a center point to obtain a preprocessed image for quality evaluation for the purpose of extracting the image quality evaluation characteristic value,

and wherein, W∈[¼*M, ⅚*M], HE [¼*N, ⅚*N], [M, N] represent the size of the original image;

after analyzing the preprocessed image for quality evaluation to extract the corresponding image quality evaluation characteristic value, the method further comprises:

normalizing each of the image quality evaluation characteristic value into a corresponding preset normalization interval.

2 . The referenceless image evaluation method of claim 1 , wherein a method for extracting the proportion fb1 of a first overexposed pixel comprises:

carrying out a grayscale processing on a color preprocessed image for quality evaluation to create a first grayscale image;

if the grayscale value of a pixel in the first grayscale image is in a preset first exposure grayscale value range, taking the current pixel as an overexposed pixel; and

taking a ratio of a total number of overexposed pixels to a total number of pixels in the first grayscale image as the proportion fb1 of the first overexposed pixels.

3 . The referenceless image evaluation method of claim 2 , wherein the method further comprises:

adjusting the value of the proportion fb1 of the first overexposed pixels to 0 if the proportion fb1 of the first overexposed pixels is smaller than a preset fourth value.

4 . The referenceless image evaluation method of claim 1 , wherein a method for extracting the proportion fb2 of a first dark pixel comprises:

carrying out a grayscale processing on a color preprocessed image for quality evaluation to create a first grayscale image;

if the grayscale value of a pixel in the first grayscale image is in a preset first dark pixel value range, taking a current pixel as a dark pixel; and

taking a ratio of a total number of dark pixels to a total number of pixels in the first grayscale image as the proportion fb2 of the first dark pixels.

5 . The referenceless image evaluation method of claim 4 , wherein the method further comprises:

adjusting the value of the proportion fb2 of the first dark pixels to 0 if the proportion fb2 of the first dark pixels is not greater than a preset fifth value.

6 . The referenceless image evaluation method of claim 1 , wherein a method for extracting the proportion fb3 of a high frequency coefficient comprises:

carrying out a grayscale processing on a color preprocessed image for quality evaluation to create a first grayscale image;

performing a Discrete Cosine Transform (DCT) on the first grayscale image to obtain the proportion fb3 of the high frequency coefficient;

i.e. fb3=length (Y<m), Y=ln(|dct(I_gray, block)|);

where, I_gray represents the first grayscale image;

dct(I_gray, block) represents performing a two-dimensional DCT on the first grayscale image I_gray with a size of block;

block=[WD, HD] represents the block size of the first grayscale image, where WD, HDϵ[2, 2{circumflex over ( )}2, 2{circumflex over ( )}3, . . . , 2{circumflex over ( )}n] on the premise of not exceeding the size of the first grayscale image;

ln represents a natural logarithm transformation with e as the base; and

length (Y<m) represents the number smaller than m in the statistic Y, and the value range of m is [−10, 0].

7 . The referenceless image evaluation method of claim 1 , wherein a method for constructing the image content evaluation model comprises:

analyzing each of the original images to extract a corresponding image content evaluation characteristic value, wherein the image content evaluation characteristic value comprises: at least one of a proportion fc1 of a non-red pixel, a proportion fc2 of a second overexposed pixel, a proportion fc3 of a second dark pixel, a number fc4 of dot impurities, and a color characteristic, wherein the color characteristic comprises at least one of a first color characteristic fc5, a second color characteristic fc6, and a third color characteristic fc7;

dividing each of the original images into a second training set and a second testing set according to a predetermined ratio, training data of the second training set by using a support vector machine, and verifying by using data of the second testing set to obtain the image content evaluation model;

wherein, both the data of the second training set and the data of the second testing set comprise the image quality calculation scores and the image content evaluation characteristic values corresponding to the original images.

8 . The referenceless image evaluation method of claim 7 , wherein before analyzing each of the original image to extract the corresponding image content evaluation characteristic value, the method further comprises:

clipping the original image at a preset size [W, H] with the center of the original image as a center point to obtain a preprocessed image for content evaluation for the purpose of extracting the image content evaluation characteristic value,

and wherein, W∈[¼*M, ⅚*M], H∈[¼*N, ⅚*N], [M, N] represent the size of the original image;

after analyzing each preprocessed image for content evaluation to extract the corresponding image content evaluation characteristic value, the method further comprises:

normalizing each of the image content evaluation characteristic value into a corresponding preset normalization interval.

9 . The referenceless image evaluation method of claim 8 , wherein a method for extracting the proportion fc1 of a non-red pixel comprises:

converting a color preprocessed image for content evaluation from RGB space to HSV space to create an HSV image;

after taking an angle metric of an H channel corresponding to each pixel in the HSV image and performing normalization, determining whether the normalized angle metric of the H channel corresponding to a current pixel is in a preset red interval; if so, marking the current pixel as 1, and if not, marking the current pixel as 0; and

taking a ratio of a total number of pixels marked as 0 to a total number of pixels in the HSV image as the proportion fc1 of non-red pixels.

10 . The referenceless image evaluation method of claim 9 , wherein the method further comprises:

adjusting the value of the proportion fc1 of non-red pixels to 0 if it is smaller than a preset sixth value.

11 . The referenceless image evaluation method of claim 8 , wherein a method for extracting the proportion fc2 of a second overexposed pixel comprises:

carrying out a grayscale processing on a color preprocessed image for content evaluation to create a second grayscale image;

if the grayscale value of a pixel in the second grayscale image is in a preset second exposure grayscale value range, taking the current pixel as an overexposed pixel; and

taking a ratio of a total number of overexposed pixels to a total number of pixels in the second grayscale image as the proportion fc2 of the second overexposed pixels.

12 . The referenceless image evaluation method of claim 11 , wherein the method further comprises:

adjusting the value of the proportion fc2 of the second overexposed pixels to 0 if the proportion fc2 statistically determined is smaller than a preset seventh value.

13 . The referenceless image evaluation method of claim 8 , wherein a method for extracting the proportion fc3 comprises:

carrying out a grayscale processing on a color preprocessed image for content evaluation to create a second grayscale image;

if the grayscale value of a pixel in the second grayscale image is in a preset second dark pixel value range, taking a current pixel as a dark pixel; and

taking a ratio of a total number of dark pixels to a total number of pixels in the second grayscale image as the proportion fc3 of the second dark pixels.

14 . The referenceless image evaluation method of claim 13 , wherein the method further comprises:

adjusting the value of the proportion fc3 of the second dark pixels to 0 if the proportion fc3 of the second dark pixels is not greater than a preset eighth value.

15 . The referenceless image evaluation method of claim 8 , wherein a method for extracting the number fc4 of dot impurities comprises:

carrying out a grayscale processing on a color preprocessed image for content evaluation to create a second grayscale image;

processing the second grayscale image using a filtering window that is a preset filtering template to form a window image;

binarizing the window image to obtain a binary image, wherein dot impurities in the binary image are assigned a value of 1, and other regions a value of 0;

counting the number of pixels with the value 1 as the number fc4 of the dot impurities.

16 . The referenceless image evaluation method of claim 15 , wherein the method further comprises:

adjusting the value of the number fc4 of the dot impurities to N if the number fc4 of the dot impurities is great than a preset ninth value, wherein the value range of Nis [0, 30];

and wherein the preset ninth value is calculated according to the values of each pixel of a R channel and a G channel in the color preprocessed image for content evaluation;

the preset ninth value thre=mean(Ir)−mean(Ig), where mean represents a mean value, Ir is the value of each pixel in the R channel, and Ig is the value of each pixel in the G channel.

17 . The referenceless image evaluation method of claim 8 , wherein a method for extracting the color characteristic comprises:

converting a color preprocessed image for content evaluation from RGB space to HSV space to create an HSV image;

obtaining the values of the R channel and the G channel in the color preprocessed image for content evaluation, and obtaining the value of an S channel in an HSV image;

then, fc5=mean(Ir)−mean(Ig),

fc6−(mean(Ir))/(mean(Ig)),

fc7=(mean (Ir))/(mean(Is));

wherein, mean represents the mean value, Ir is the value of each pixel in the R channel, Ig is the value of each pixel in the G channel, and Is is the value of each pixel in the S channel.

18 . The referenceless image evaluation method of claim 1 , wherein before constructing the image quality evaluation model and the image content evaluation model, the method further comprises:

performing initial scoring on the m original images using n groups of rules to generate m*n groups of evaluation scores;

standardizing the m*n groups of evaluation scores to obtain m*n groups of standard scores x mn ′;

x mn ′=(x mn −μ m )/σ m , where, x mn represents an initial score of any original image calculated by any rule; μ m represents a mean value of m initial scores obtained from m original images according to the rules corresponding to x mn ;

σ m represents a variance of m initial scores obtained from m original images according to the rules corresponding to x mn;

eliminating abnormal standard scores from the m*n groups of evaluation scores while keeping effective standard scores; wherein

if (x mn ′−μ n )/σ n >score, score≥μ n −3σ n , the current standard score is determined as an abnormal value; if (x mn ′−μ n )/σ n ≤score, the current standard score is determined as an effective value;

where, μ n represents a mean value of n initial scores obtained from the original images corresponding to x mn ′ using n groups of rules; On represents a variance of n initial scores obtained from the original images corresponding to x mn ′ using n groups of rules;

taking a mean value, a median value or a weighted value of the effective standard scores corresponding to each original image as an evaluation score of the current original image, wherein the evaluation score comprises: the image quality calculation score or the image content calculation score.