CN107977976A - A kind of image partition method - Google Patents

Abstract

The invention discloses an image segmentation method. The image has object pixels and background pixels. The image segmentation method includes: (1) taking an image to be segmented, deleting the background, and then performing binary processing on the image; (2) ) calculate the Euclidean distance from each item pixel to the nearest background pixel, and assign a value to the item pixel with the Euclidean distance; (3) segment the image by using the watershed algorithm to form multiple regions; (4) use The support vector machine identifies the number of samples in each region. If the number of samples is greater than 1, the region is divided again by the watershed algorithm. The present invention assigns a value to the item pixel by the Euclidean distance from the item pixel to the nearest background pixel, then uses the watershed algorithm to segment the image, and uses the support vector machine to identify the number of samples in the segmented area, and if the sample has adhesion, it will be segmented again. Existing segmentation methods are able to separate objects with severe adhesions or objects whose shape will change after adhesions.

Description

A Method of Image Segmentation

technical field

The invention relates to an image processing method, in particular to an image segmentation method.

Background technique

Image segmentation is the technology and process of dividing an image into several specific regions with unique properties and proposing objects of interest. It is a key step from image processing to image analysis. The objects extracted after image segmentation can be used in image semantic recognition, image search and other fields.

At present, the main image segmentation methods include threshold segmentation, region segmentation, edge segmentation, etc. The research objects mainly focus on the segmentation of items with regular shapes and not easy to stick and squeeze, such as mechanical parts, fresh fruits, rice, etc. For example, Chinese patent CN101556693 discloses a watershed SAR image segmentation method for threshold method marker extraction, including the following steps: (1) Gaussian low-pass filtering is carried out to the original image Img to obtain the filtered image GImg; (2) the filtered image GImg is obtained. Gradient to obtain the gradient image PGImg; (3) extract the internal label LImg from the original image Img by Otsu method; (4) perform classical watershed transformation on the internal label to obtain the external label WLImg; (5) use the internal and external labels to gradient the gradient image PGImg fix. Use the forced minimum technique to correct, so that the local minimum region only appears in the marked position; (6) Perform watershed transformation on the corrected gradient map, and the resulting map is the final segmentation map RImg. This patent combines the OTSU threshold method with the marked watershed, and uses the thresholded image as the source of the marked watershed mark, which effectively eliminates the influence of texture information.

However, there are relatively few studies on the segmentation of some irregular and deformable object samples, such as soft and deformable dried fruits such as raisins. The shape and size of each of these samples are different, and the color is the same, and there will be mutual adhesion and extrusion deformation between two adjacent samples, which increases the difficulty of segmentation.

Contents of the invention

The invention provides an image segmentation method, which can effectively segment the soft and deformable items in the image, accurately calculate the number of item samples, and is less disturbed by the extrusion deformation of the items.

An image segmentation method, the image has item pixels and background pixels, the image segmentation method comprising:

(1) Get the image to be segmented, delete the background, and then perform binarization on the image;

(2) Calculate the Euclidean distance from each item pixel to the nearest background pixel, and assign a value to the item pixel with the Euclidean distance;

(3) using a watershed algorithm to segment the image to form multiple regions;

(4) Use the support vector machine to identify the number of samples in each region. If the number of samples is greater than 1, then use the watershed algorithm to segment the region again.

By calculating the Euclidean distance from the item pixel to the nearest background pixel, and assigning the item pixel with the Euclidean distance, the image gradient can be formed, and the watershed algorithm can be used for segmentation.

The samples refer to the number of items in the image, and if the number of samples is greater than 1, it means that the segmentation is not complete and needs to be segmented again, and the image depth gradient h needs to be changed when segmenting again.

Before noise reduction, there will be noise points in the background and samples, which will affect the segmentation result. Preferably, the image is denoised after the background is deleted.

Preferably, in step (2), the image depth value gradient of the watershed algorithm h=4 and adopts 8-connectivity. If the depth gradient is too large, the proper low-lying land will be lost. If the depth gradient is too small, too many low-lying lands will be identified, resulting in over-segmentation. Here, choosing 4 can get better results; using 8-connectivity can reduce the number of sided objects. Loss of edge pixels.

Preferably, the kernel function of the support vector machine adopts radial odd function, degree d=4, kernel coefficient γ=1/n, tolerance tol=10 ^-6 , and error parameter C=1.0. Compared with other functions (such as the sigmoid function), the radial basis function is less prone to overfitting.

Preferably, the input variables of the support vector machine are the eccentricity, area, abscissa of the centroid, ordinate of the centroid, major axis, minor axis and perimeter of the sample; the output variable is the number of samples. The above seven parameters can effectively describe the shape characteristics of the measured object.

Preferably, in step (3), the image depth value gradient of the watershed algorithm h=3 and adopts 8-connectivity. Preferably, step (4) is performed after repeating steps (2) and (3) several times in sequence.

Preferably, in the binarization process, item pixels are set to 1, and background pixels are set to 0.

The image is a tiled image of cohesive raisins.

Preferably, in step (3), before the image is segmented, the local minimum is filtered to avoid over-segmentation of the image.

The present invention assigns a value to the item pixel by the Euclidean distance from the item pixel to the nearest background pixel, then uses the watershed algorithm to segment the image, and then uses the support vector machine to identify the number of samples in the segmented area, and if the sample has adhesion, then segment again. Because the existing segmentation methods can separate objects with serious adhesion or changes in shape after adhesion.

Description of drawings

Figure 1 shows the RGB image to be segmented.

Figure 2 is the image after background removal and noise reduction preprocessing.

Figure 3 is the result map of the item pixel assignment Euclidean distance.

Figure 4 is the result of image segmentation using the watershed algorithm.

Figure 5a is a distribution map of local minima.

Fig. 5b is a segmentation result diagram after filtering the local minimum.

Fig. 6 is the segmentation result of the part whose number of samples is greater than 1 identified by the support vector machine. Fig. 6a is the segmentation result figure of the part represented by the 7th line in Table 1, Fig. 6b is the segmentation result figure of the part represented by the 17th line in Table 1, and Fig. 6c is the part represented by the 31st line in Table 1 Segmentation result graph.

FIG. 7 is a graph of the prediction results of the support vector machine for each part of the sample size.

Figure 8 is an overall rendering of the segmentation.

Detailed ways

Taking the raisins of rose purple variety as an example, the raisins RGB image is collected (as shown in Figure 1). The raisins are basically in a flat state, but some adjacent raisins overlap and stick together, because the color difference and shape difference between the raisins are not obvious, and the raisins It is relatively soft and prone to deformation, and cannot be divided well by traditional methods.

The method of the present invention firstly adopts difference method to delete the background of the image, then adopts mean filtering to carry out noise filtering on the image, then carries out binarization processing on the image, obtains the binary image as shown in Figure 2, and the pixel with a value of 1 ( White) are raisin pixels, and pixels with a value of 0 (black) are background pixels.

Then calculate the Euclidean distance of each raisin pixel in the image. For each pixel with a value of 1, find the closest zero pixel to it, calculate the Euclidean distance between two pixels according to the following formula, and assign the calculation result to the original pixel, the result is shown in Figure 3. The formula for calculating the Euclidean distance is:

(2) Use the H minimum value transformation function to transform the gradient of the image depth value to 4, and use the 8-connected watershed algorithm to operate on the image, and the results are shown in Figure 4. In order to solve the over-segmentation problem of the watershed algorithm, the method of filtering local minimum values is adopted, and two local minimum values are selected for each watershed partition. The results are shown in Figure 5.

(3) The support vector machine is used to identify the number of samples of each divided part, and the part with the number of samples greater than 1 is extracted separately. The kernel function of the support vector machine adopts the radial odd function, the degree d=4, the kernel coefficient γ=1/n, the tolerance tol=10 ⁻⁶ , and the error parameter C=1.0. The input variables of the support vector machine are seven eigenvalues of the sample, including eccentricity, area, centroid abscissa, centroid ordinate, major axis, minor axis, and perimeter, and the output variable is the sample size. Table 1 shows the input variable values, predicted values, and reference values of the support vector machine test set. Figure 7 labels the eccentricity of each segment that was segmented. It can be seen from Table 1 and Figure 7 that the support vector machine accurately identified the number of samples corresponding to rows 7, 17, and 31 in Table 1 as 2, and there were no other misidentified items.

Table 1. Input values, predicted values, and reference values of the support vector machine test set

The accuracy of the above prediction results is determined by the three indicators of sensitivity, accuracy and specificity, and the three indicators are calculated by formula 2, formula 3 and formula 4 respectively. where a is the number of true negatives, b is the number of false positives, c is the number of false negatives, and d is the number of true positives. A positive reference value is 2, and a negative reference value is 1. Table 2 shows the recognition accuracy metrics.

Sensitivity = d/(c+d) Formula 2

Accuracy = d/(b+d) Formula 3

Specificity = a/(a+b) Formula 4

Table 2. Accuracy of sample size identification

(4) Use the watershed algorithm to segment the part whose number of samples is greater than 1, that is, the part represented by rows 7, 17, and 31 in Table 1. The segmentation results are shown in Figure 6. The overall segmentation result rendering is shown in Figure 8.

Claims (10)

1. An image segmentation method, the image has article pixels and background pixels, and the image segmentation method comprises: (1) Get the image to be segmented, delete the background, and then perform binarization on the image; (2) Calculate the Euclidean distance between each item pixel and the nearest background pixel, and assign a value to the item pixel with the Euclidean distance; (3) using a watershed algorithm to segment the image to form multiple regions; (4) Use the support vector machine to identify the number of samples in each region. If the number of samples is greater than 1, then use the watershed algorithm to segment the region again. 2. image segmentation method as claimed in claim 1 is characterized in that, image is carried out denoising processing after deleting background. 3. The image segmentation method according to claim 1, characterized in that, in step (2), the image depth value gradient h=4 of the watershed algorithm adopts 8-connectivity. 4. image segmentation method as claimed in claim 1, is characterized in that, the kernel function of described support vector machine adopts radial odd function, degree d=4, kernel coefficient γ=1/n, tolerance tol= ^{10- 6} , error parameter C=1.0. 5. image segmentation method as claimed in claim 1, is characterized in that, the input variable of described support vector machine is eccentricity, area, centroid abscissa, centroid ordinate, major axis, minor axis, circumference of sample; The output variable is the sample size. 6 . The image segmentation method according to claim 1 , wherein in step (3), the image depth value gradient h=3 of the watershed algorithm is 8-connected. 7. The image segmentation method according to claim 1, characterized in that, after repeating steps (2) and (3) several times in sequence, step (4) is then carried out. 8 . The image segmentation method according to claim 1 , wherein, in the binarization process, object pixels are set to 1, and background pixels are set to 0. 9 . 9. The image segmentation method according to claim 1, characterized in that, the image is a tiled and cohesive raisin image. 10. The image segmentation method according to claim 1, wherein in step (3), before the image segmentation, local minimum values are filtered.