Segmentation system for intestinal fecal water area image

Abstract

The invention belongs to the technical field of image processing, and provides a segmentation system for an intestinal fecal water area image, which comprises the following steps of: firstly, performing local denoising by combining median filtering and bilateral filtering for an intestinal preparation image, then performing overall denoising by using a non-local average denoising algorithm, and taking an image block as a unit; similar areas are found in the intestinal tract preparation image, Gaussian noise in the image can be well removed, a parameter is added to stipulate the number of image blocks needing to be calculated in the matching process, different images can be better adapted, the problem that when the number of the image blocks needing to be calculated is too large, the influence of noise and artifacts can be reduced is solved, and the image matching accuracy is improved. According to the method, the problems that the number of image blocks needs to be calculated, but the details of the image are reduced, more details can be reserved when the number of the image blocks needs to be calculated is too small, but the influence of noise and artifacts is increased are solved, edge feature enhancement detection is combined, and the problem that edge information and the like are lost is solved on the basis that the denoising effect is guaranteed.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and particularly relates to a segmentation system for images of intestinal fecal water regions. Background Art

[0002] The segmentation of the fecal water region image of the intestinal preparation image is a method for automatically judging whether the intestine is cleaned after the intestine is cleaned before intestinal surgery. During the acquisition or transmission of the intestinal preparation image, it is usually affected by the performance of the hardware itself or environmental factors such as light and temperature, which greatly reduces the quality of the image and affects subsequent higher-level image analysis. Therefore, reducing the influence of noise and edge information is very important for obtaining the segmentation image of the fecal water region of the intestinal preparation image.

[0003] The inventor found that when processing the image of the intestinal fecal water region, threshold segmentation is a simple region-based segmentation technology and also a basic image segmentation method. This segmentation method assumes that the gray levels in each region are uniform, but the gray levels between adjacent regions are different; since the threshold segmentation method only considers the pixel values of the target and the background, and does not consider other features in the image of the intestinal fecal water region, such as spatial features, color and texture features, this method is very sensitive to noise. Due to the existence of Gaussian noise, filtering only the local regions and some pixel points of the image cannot achieve a good denoising effect. Therefore, the non-local means denoising (NLM) algorithm is adopted. However, when the NLM algorithm is adopted, the existence of different noises in the image will interfere with some similar pixel points, resulting in inaccurate calculation of weights and the phenomenon of edge information loss in the image. Summary of the Invention

[0004] In order to solve the above problems, the present invention proposes a segmentation system for images of intestinal fecal water regions, which first performs local denoising and then performs overall denoising, and searches for similar regions in the image in units of image blocks, and can better remove the Gaussian noise existing in the image.

[0005] In order to achieve the above object, the present invention provides a segmentation system for images of intestinal fecal water regions, and adopts the following technical solutions:

[0006] A segmentation system for images of intestinal fecal water regions, comprising:

[0007] A data acquisition module, configured to: acquire an intestinal preparation image;

[0008] A denoising module, configured to: for the bowel preparation image, first perform local denoising by combining median filtering and bilateral filtering, and then perform global denoising using a non-local mean denoising algorithm; during global denoising, taking image patches as units, searching for similar regions in the bowel preparation image, and adding a parameter to specify the number of image patches to be calculated during the matching process;

[0009] An edge feature enhancement module, configured to: perform a discrete Fourier transform on the globally denoised image to obtain a denoised image; perform edge detection on the obtained denoised image to enhance edge features;

[0010] An image segmentation module, configured to: segment the denoised image with enhanced edge features.

[0011] Further, during median filtering, the median value of the pixel values in a neighborhood of a to-be-processed pixel in the bowel preparation image is selected to replace the to-be-processed pixel, thereby eliminating isolated noise points.

[0012] Further, the pixel values in a neighborhood of a to-be-processed pixel are arranged in a sequence in ascending or descending order, and the value at the middle position in the sequence is the median value.

[0013] Further, during bilateral filtering, the spatial domain weight and the gray domain weight are non-linearly combined to form a new weight value and a new filtering template, and the new filtering template is convolved with the noisy image for denoising.

[0014] Further, adding a parameter as a group size parameter, where the group size parameter specifies the size of the group when a reference image patch is compared with other image patches; when determining the parameter, consider the illumination and temperature during the acquisition of the bowel preparation image, as well as the median value, and fine-tune the number of patches considered when calculating the similarity weight; specifically, calculate the ratios of the illumination to a preset illumination value, the temperature to a preset temperature value, and the median value to a preset median value respectively, perform a weighted sum on the three ratios to obtain an adjustment parameter, compare the obtained adjustment parameter with a preset parameter, when the adjustment parameter is larger than the preset parameter, for each increase of a preset interval in the difference between the adjustment parameter and the preset parameter, the number of patches considered when calculating the similarity weight increases by a preset number; when the adjustment parameter is smaller than the preset parameter, for each increase of a preset interval in the absolute value of the difference between the adjustment parameter and the preset parameter, the number of patches considered when calculating the similarity weight decreases by a preset number.

[0015] Further, after denoising using the non-local mean denoising algorithm, the weight of the weighted sum of a pixel in the image and the pixels in the image neighborhood is equal to the similarity between the pixels.

[0016] Further, based on the discrete Fourier transform, adding a window function for weighting or scaling the time-domain signal.

[0017] Further, initialize the overall denoised image and perform a discrete Fourier transform on the initialized image.

[0018] Further, when performing edge detection, add a convolution kernel size parameter to adapt to different images and application scenarios; adjust the convolution kernel size parameter according to the characteristics of the image, and the convolution kernel size parameter determines the size of the convolution kernel.

[0019] Further, use the two-dimensional maximum between-class variance method for multiple processes to make the probabilities of gray pixels in the processed edge information area and noise area less than the set probability value. Curve fit the obtained several two-dimensional segmentation threshold points, and perform binary processing on each gray pixel according to the segmentation threshold curve to achieve image segmentation.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0021] 1. For the bowel preparation image in the present invention, first perform local denoising by combining median filtering and bilateral filtering, and then use the non-local mean denoising algorithm for overall denoising. Taking image blocks as units, search for similar regions in the bowel preparation image, which can better remove the Gaussian noise existing in the image. Adding a parameter to specify the number of image blocks to be calculated during the matching process can better adapt to different images, solve the problem that when the number of image blocks to be calculated is too large, the influence of noise and artifacts can be reduced, but the details of the image will also be reduced, and when the number of image blocks to be calculated is too small, more details can be retained, but the influence of noise and artifacts is increased. Combining edge feature enhancement detection can solve the problem of loss of edge information and the like on the basis of ensuring the denoising effect;

[0022] 2. The present invention adds a wave characteristic term on the basis of the previous diffusion equation, making the model between the characteristics of the diffusion and wave equations. Therefore, in terms of maintaining the edge structure and texture of the image, it is superior to the diffusion equation model;

[0023] 3. The present invention removes the influence of noise in the image and well maintains the real edge structure and important texture details in the image;

[0024] 5. The method for segmenting the fecal water area image of the bowel preparation image in the present invention can separate the useful information from the edge information area and the noise area to the greatest extent, effectively reducing the influence of edge information or noise on the image segmentation process, and making the segmentation effect more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The accompanying drawings forming a part of this embodiment are used to provide a further understanding of this embodiment. The schematic embodiments and descriptions thereof of this embodiment are used to explain this embodiment and do not constitute an improper limitation of this embodiment.

[0026] Figure 1 This is the flowchart of Embodiment 1 of the present invention.

[0027] Figure 2 This is the image collected in Embodiment 1 of the present invention;

[0028] Figure 3 This is the image after denoising in Embodiment 1 of the present invention;

[0029] Figure 4 This is the image after edge feature enhancement - discrete Fourier transform in Embodiment 1 of the present invention;

[0030] Figure 5 This is the image after Sobel edge detection in Embodiment 1 of the present invention;

[0031] Figure 6 This is the image segmentation result in Embodiment 1 of the present invention. Detailed implementation manners

[0032] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0033] It should be noted that the following detailed description is exemplary and is intended to provide further illustration of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present application belongs.

[0034] Embodiment 1:

[0035] Due to the existence of Gaussian noise, filtering only local regions and some pixel points of the image cannot achieve good denoising effects. Therefore, the non-local means denoising (NLM) algorithm is adopted. However, when using the NLM algorithm, the existence of different noises in the image will interfere with some similar pixel points, resulting in inaccurate calculation of weights and easy loss of edge information in the image;

[0036] As Figure 1 shown, this embodiment provides a segmentation system for images of the intestinal fecal water area, including:

[0037] A data acquisition module, configured to: acquire intestinal preparation images;

[0038] A denoising module, configured to: for the intestinal preparation image, first perform local denoising by combining median filtering and bilateral filtering, and then perform overall denoising using the non-local means denoising algorithm; during overall denoising, taking image blocks as units, searching for similar regions in the intestinal preparation image, and adding a parameter to specify the number of image blocks to be calculated during the matching process;

[0039] An edge feature enhancement module, configured to: perform a discrete Fourier transform on the overall denoised image to obtain a denoised image; perform edge detection on the obtained denoised image to enhance edge features;

[0040] An image segmentation module, configured to: segment the denoised image with enhanced edge features.

[0041] Specifically, for the bowel preparation image, first perform local denoising by combining median filtering and bilateral filtering, and then perform overall denoising using the non-local mean denoising algorithm. Taking image blocks as units, search for similar regions in the bowel preparation image, which can better remove the Gaussian noise existing in the image. Adding a parameter to specify the number of image blocks to be calculated during the matching process can better adapt to different images, solving the problem that when the number of image blocks to be calculated is too large, it can reduce the influence of noise and artifacts but also reduce the details of the image, and the problem that when the number of image blocks to be calculated is too small, it can retain more details but increases the influence of noise and artifacts. Combining edge feature enhancement detection solves the problem of loss of edge information, etc. on the basis of ensuring the denoising effect. It can be understood that the bowel preparation image in this embodiment refers to the fecal water image taken after the patient cleans the bowel through specific preparation measures before certain medical examinations or treatments, so as to ensure that doctors can accurately observe the internal structure of the bowel during the examination or treatment process.

[0042] The specific content implemented by each module in this embodiment is as follows:

[0043] S1. Obtain a bowel preparation image / a bowel fecal water region image to be processed, and use median filtering. By selecting the median of the pixel values in a neighborhood of a pixel to be processed in the bowel preparation image, replace the pixel to be processed, thereby eliminating isolated noise points. Specifically, first, arrange the pixel values in a neighborhood of a pixel to be processed in a sequence in ascending or descending order, and the value at the middle position of the sequence is the required median; secondly, replace the pixel value to be processed with the median to make the pixel values in the neighborhood closer to the true values.

[0044] Next, use a bilateral filter. On the basis of a Gaussian filter, considering the similarity of pixel gray values in the filter template, non-linearly combine the spatial domain weight and the gray domain weight to form a new weight value, and perform a convolution operation on the new filter template and the noisy image to obtain a denoised image.

[0045] Due to the existence of Gaussian noise, filtering only the local regions and some pixel points of the image cannot achieve a good denoising effect. When comparing the similarity of images, the similarity of gray values between individual pixel points is no longer used as the comparison criterion. Therefore, it is necessary to calculate the neighborhoods of pixels, and these neighborhoods are the search regions of the NLM algorithm. For the NLM algorithm, in this embodiment, an algorithm calculation group size is added to better adapt to different images; the group size specifies the size of the group when the reference patch is compared with other patches, that is, the group size specifies the number of patches to be calculated during the template matching process; generally, a smaller group size will increase the influence of noise and artifacts, but more details can be retained; on the contrary, a larger group size will reduce the influence of noise and artifacts, but also reduce the details of the image; in this embodiment, first, the estimate_sigma() function is used to calculate the noise standard deviation of the image; then, the patch_size, patch_distance, group_size, and h parameters are set; among them, group_size specifies the group size parameter, which specifies the number of patches considered when calculating the similarity weights. patch_size is the size of the patch used to calculate the similarity weights in the NLM algorithm, which determines the size of each patch, usually a square region; h is the smoothing parameter in the NLM algorithm. It controls the degree of denoising, and by adjusting the value of h, the balance between the denoising effect and the retention of image details can be achieved; finally, in this embodiment, the previously calculated parameters and the image are used as inputs, and the denoise_nl_means() function is used to apply the NLM algorithm to obtain the denoised image.

[0046] In another embodiment, when the parameter group_size is determined, the illumination, temperature, and median size during the acquisition of bowel preparation images are considered, and the number of patches considered when calculating the similarity weights is fine-tuned; specifically, the ratios of illumination to the preset illumination value, temperature to the preset temperature value, and median to the preset median are calculated respectively, and the three ratios are weighted and summed to obtain an adjustment parameter. The weights can be optionally 0.3, 0.2, and 0.5. The obtained adjustment parameter is compared with the preset parameter. When the adjustment parameter is larger than the preset parameter, it indicates that the detail degree during image acquisition is better. At this time, for each increase in the preset interval of the difference between the adjustment parameter and the preset parameter, the number of patches considered when calculating the similarity weights increases by a preset number, reducing the influence of noise and artifacts while retaining image details; when the adjustment parameter is smaller than the preset parameter, it indicates that the detail degree during image acquisition is poorer. At this time, for each increase in the preset interval of the absolute value of the difference between the adjustment parameter and the preset parameter, the number of patches considered when calculating the similarity weights decreases by a preset number, retaining image details while reducing the influence of noise and artifacts.

[0047] The NLM algorithm can be expressed by Equation (1):

[0048]

[0049] Among them, f(x) is the given discrete noise image; is the estimated value, which is the weighted average of all pixels in the image; λ(x) is the neighborhood of pixel x; ω(x, y) is the weight of the original image f(x), and 0 < ω(x, y) < 1, ∑ y∈λ(x) ω(x, y) = 1, x ∈ λ(x), y ∈ λ(x); The square of the difference in luminance values between pixel x and pixel y is the pixel similarity, and the similarity between pixels is measured by the Euclidean distance between vectors within the neighborhood. The gray-level neighborhood similar to the intensity gray-level vector has a larger weight on average, and these weights are also called Euclidean distances, which can be defined as Equation (2):

[0050]

[0051] Among them, k(x) is the sum of all weights and is also a normalization factor; a is a constant; λ(x) and λ(y) are the neighborhoods of pixel x and pixel y respectively; is the Gaussian-weighted Euclidean distance between the neighborhood of pixel x and the neighborhood of pixel y; δ 2 is the filtering coefficient. After each weight in the image is divided by k(x), the weight is greater than 0 and the sum of the weights is 1. When δ > 0, the degree of filtering is low, and the exponential function decays, and the weight of the Euclidean distance function decreases. The block (Patch) neighborhood is the neighborhood of pixel and pixel, and generally it is smaller than the search area. When calculating the Euclidean distance, the closer to the center of the Patch neighborhood, the greater the weight, and vice versa, and the weight follows a Gaussian distribution. In practice, a uniform distribution can also be used to reduce the computational complexity. After denoising using the NLM algorithm, the weight of the weighted sum of the pixel in the image and the pixels in the image neighborhood is equal to the similarity between pixel x and pixel y. Defining a decreasing function of the weighted Euclidean distance represents the similarity between pixel x and pixel y, and the intensity gray-level vector affects the weighted Euclidean distance. In the noisy neighborhood, the calculation of the Euclidean distance is shown in Equation (3):

[0052]

[0053] Among them, represents the Gaussian-weighted Euclidean distance between the neighborhood of pixel x and the neighborhood of pixel y, where a > 0 is the standard deviation of the Gaussian kernel. Optionally, α = 1.8.

[0054] S2. Initialize the obtained image and its parameters:

[0055] Set the time step τ, the number of stopping iterations L, the initial iteration number n = 1, and initialize the image Among them, For the original image u o The expanded image; based on the discrete Fourier transform, a window function parameter can be added to better adapt to certain signals. The window function is a function that weights or scales the time-domain signal, so that the interference of the signal at certain frequencies can be selectively reduced or eliminated. Next, the initialized image is subjected to the discrete Fourier transform to obtain and Iteratively calculate and solve the discretized space fractional diffusion-wave equation model (4) through formula (5):

[0056]

[0057] where the space fractional diffusion-wave equation model is denoted as u(p, t), t ∈ T, Ω is the two-dimensional image domain, T is a positive time constant; λ > 0 is the weight parameter, optionally, λ = 7; u o represents the image smoothed by a Gaussian kernel with a standard deviation of σ; g is the diffusion function used to adaptively adjust the diffusion intensity according to the local features of the image, and is defined here as g(s) = 1 / (1 + (s / k) 2 ), k > 0; uo(x) is the initial noisy image; the fractional divergence operator and the fractional gradient operator are respectively defined as and The order satisfies 1 ≤ α ≤ 2; where τ is the time step, optionally, τ = 0.2, n is the number of iterations, optionally, the number of iterations is 400, ^ represents the Fourier transform of the corresponding variable, M n is the spatial difference approximation of in formula (4); the iterative result is subjected to the inverse discrete Fourier transform to obtain the final result u n+1 , u n+1 is the denoised image.

[0058] S3. Sobel edge detection enhances the boundary features in an image by detecting the differences in image brightness and determines the edge information of the image. In image processing, the image edges are described by amplitude and direction attributes. When the convolution kernel size is 3x3, the Sobel operator can detect finer details and smaller edges, but there may be noise and discontinuous edges; when the convolution kernel size is 5×5, the Sobel operator can detect larger edges and regions, but some image detail information may be lost. Based on the Sobel operator, in this embodiment, a parameter named convolution kernel size is added to better adapt to different images and application scenarios. The convolution kernel size determines the size of the Sobel operator's convolution kernel, usually 3×3 or 5×5; therefore, the convolution kernel size can be adjusted according to the characteristics of the image. Specifically, first read the image and set the convolution kernel size, then use the cv2.Sobel() function to apply the Sobel operator for edge detection, and finally, calculate the amplitude and direction and use the cv2.imshow() function to display the original image and the Sobel edge detection result. By adjusting the convolution kernel size, the Sobel operator can better adapt to different images and filtering requirements, thus improving the edge detection result. The amplitude and direction of the Sobel operator are represented by equations (6) and (7) respectively:

[0059]

[0060] where g a is the gradient component in the horizontal direction; g b is the gradient component in the vertical direction.

[0061] S4. Perform grayscale processing on the denoised image to obtain a grayscale image; acquire the grayscale and neighborhood average grayscale of each pixel point in the grayscale image, and establish a coordinate system; use the two-dimensional maximum between-class variance method to process the pixels of the grayscale image under the coordinate system to determine the two-dimensional segmentation threshold point; the two-dimensional segmentation threshold point divides the two-dimensional grayscale histogram of the original image into four regions, namely the background region, the target region, the edge information region, and the noise region; judge the probability that the number of pixels in the edge information region and the noise region accounts for the total number of pixels. When the probability is greater than or equal to the set probability value, use the two-dimensional maximum between-class variance method to process the pixels of the grayscale image in the edge information region and the noise region respectively, and re-determine the two-dimensional segmentation threshold points corresponding to the edge information region and the noise region. The corresponding two-dimensional segmentation threshold points respectively re-divide the edge information region and the noise region into four regions: until the probability that the number of pixels in the re-divided edge information region and the noise region accounts for the total number of pixels is less than the probability value; analyze the determined two-dimensional segmentation threshold points. Next, use the Fourier parameter model to describe the curve. According to Bayes' theorem, give an objective function according to the principle of maximum a posteriori probability, and determine the Fourier coefficients by maximizing the objective function. According to the segmentation experience of similar images, give an initial curve, and then optimize the objective function according to the image data in a specific segmentation example to change the parameters of the initial curve, fit the image data, and obtain a specific curve determined by the image data. Optionally, use the least squares method to perform curve fitting on each segmentation threshold point. Obtain the segmentation threshold curve, and perform binarization processing on each pixel point according to the segmentation threshold curve to achieve image segmentation; optionally, performing binarization processing on each pixel point according to the segmentation threshold curve includes: comparing whether the neighborhood average grayscale value corresponding to each pixel point of the segmentation threshold curve is greater than or equal to the neighborhood average grayscale value obtained by performing neighborhood calculation on the pixel. If it is greater than or equal to, set the pixel value to the preset grayscale value of the target region; if it is not greater than or not equal to, set the pixel value to the preset grayscale value of the background region; the segmentation threshold curve can be a linear function curve, a quadratic function curve, or a cubic function curve. The value range of the probability value can be [0.01, 0.05].

[0062] The above are only the preferred embodiments of this embodiment and are not used to limit this embodiment. For those skilled in the art, this embodiment can have various changes and modifications. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of this embodiment shall be included within the protection scope of this embodiment.

Claims

1. A segmentation system for intestinal feces and water area images, characterized in that: include: The data acquisition module is configured to: acquire a bowel preparation image; The denoising module is configured to: for the intestinal preparation image, firstly perform local denoising by combining median filtering and bilateral filtering, and then perform overall denoising by using a non-local average denoising algorithm; When denoising overall, similar regions are searched in the intestinal preparation images in units of image blocks, and a parameter is added to specify the number of image blocks to be calculated during the matching process; The edge feature enhancement module is configured to: perform discrete Fourier transform on the overall denoised image to obtain a denoised image; perform edge detection on the obtained denoised image to enhance edge features; The image segmentation module is configured to segment the denoised image after enhancing edge features.

2. A segmentation system for intestinal feces and water area images as claimed in claim 1, characterized in that: During median filtering, the pixel to be processed is replaced by the median value of each pixel in a neighborhood of the pixel to be processed in the intestinal preparation image, thereby eliminating isolated noise points.

3. A segmentation system for intestinal feces and water area images as claimed in claim 2, characterized in that: The pixel values ​​in a neighborhood of the pixel to be processed are arranged in a sequence from small to large or from large to small, and the value in the middle of the sequence is the median.

4. A segmentation system for intestinal feces and water area images as claimed in claim 1, characterized in that: During bilateral filtering, the spatial domain weights and grayscale domain weights are nonlinearly combined to form new weight values ​​and a new filter template, and the new filter template is convolved with the noisy image for denoising.

5. A segmentation system for intestinal feces and water area images as claimed in claim 1, characterized in that: A parameter is added as a group size parameter, and the group size parameter specifies the size of the group when the reference image block is compared with other image blocks; when the parameter is determined, the illumination and temperature when the intestinal preparation image is obtained, as well as the median size, are considered to fine-tune the number of blocks considered when calculating the similarity weight; specifically, the ratios of the illumination to the preset illumination value, the temperature to the preset temperature value, and the median to the preset median are calculated respectively, and the three ratios are weighted summed to obtain an adjustment parameter, and the obtained adjustment parameter is compared with the preset parameter. When the adjustment parameter is larger than the preset parameter, the number of blocks considered when calculating the similarity weight increases by a preset number for each preset interval of the difference between the adjustment parameter and the preset parameter; When the adjustment parameter is smaller than the preset parameter, the number of blocks considered when calculating the similarity weight is reduced by a preset number for each increase in the absolute value of the difference between the adjustment parameter and the preset parameter by a preset interval.

6. A segmentation system for intestinal feces and water area images as claimed in claim 1, characterized in that: After denoising using the non-local mean denoising algorithm, the weight of the weighted sum of the pixels in the image and the pixels in the image neighborhood is equal to the similarity between the pixels.

7. A segmentation system for intestinal feces and water area images as claimed in claim 1, characterized in that: On the basis of discrete Fourier transform, a window function is added to weight or scale the time domain signal.

8. A segmentation system for intestinal feces and water area images as claimed in claim 1, characterized in that: The overall denoised image is initialized, and discrete Fourier transform is performed on the initialized image.

9. A segmentation system for intestinal feces and water area images as claimed in claim 1, characterized in that: When performing edge detection, add a convolution kernel size parameter to adapt to different images and application scenarios; adjust the convolution kernel size according to the characteristics of the image.

10. The segmentation system for intestinal feces and water area images according to claim 1, characterized in that: The two-dimensional maximum inter-class variance method is used for multiple processing to make the probability of grayscale pixels in the edge information area and the noise area after processing less than the set probability value. The processed two-dimensional segmentation threshold points are curve fitted, and each grayscale pixel is binarized according to the segmentation threshold curve to achieve image segmentation.

Images