CN111583279A - A Superpixel Image Segmentation Method Based on PCBA - Google Patents

Abstract

The invention discloses a superpixel image segmentation method based on PCBA, comprising: obtaining an original image; improving the contrast of the original image to obtain an input image; calculating the saliency value of the input image and generating a saliency map; performing superpixel segmentation on the saliency map, and Obtain the superpixel label; create a superpixel image based on the superpixel label, the pixel value of each pixel in the superpixel image is the average gray value of the corresponding superpixel block; divide the original image into the target area and the background based on the optimal segmentation threshold region, and generate a segmentation map consisting of the target region and the background region. Aiming at the characteristics that the color information of PCBA is simple and does not contain complex background information, the present invention uses superpixel labels to generate superpixel images before final segmentation. The omission of the target reduces the pixels of the image to be processed, thereby reducing the computational cost and improving the efficiency of segmentation.

Description

A Superpixel Image Segmentation Method Based on PCBA

technical field

The invention belongs to the field of image processing, and in particular relates to a PCBA-based superpixel image segmentation method.

Background technique

Superpixels belong to image segmentation technology, which refers to irregular pixel blocks with certain visual significance composed of adjacent pixels with similar texture, color, brightness and other characteristics. The superpixel segmentation algorithm is an important preprocessing tool that reduces computation and is widely used in object tracking, object recognition, 3D reconstruction, image segmentation, and saliency feature extraction. Compared with pixels, superpixels can remove redundant information and improve the operation speed. It can be used as a common method of preprocessing in saliency detection, which can greatly improve the operation efficiency while saving edge information as much as possible. At present, superpixel algorithms can be simply divided into gradient descent-based algorithms and clustering-based algorithms.

Both the watershed algorithm and the Mean-Shift algorithm are classic gradient descent algorithms. The watershed algorithm uses gradient descent to obtain the similarity between pixels to generate a closed contour. The Mean-Shift algorithm is an iterative process, looking for the average moving vector of the current pixel, moving to the vector position and continuing to search until a certain condition is met. Although the watershed algorithm has a faster operation speed, it is easy to over-segment, and Mean-Shift is the opposite. K-means clustering (K-means) algorithm, Linear Spectral Clustering (LSC) algorithm, Simple Linear Iterative Clustering (SLIC) algorithm and Simple Non-Iterative Clustering (Simple Non-Iterative Clustering) algorithm are all Classic clustering algorithm. The LSC algorithm, the SLIC algorithm and the SNIC algorithm are all improvements based on the K-means algorithm. The K-means algorithm randomly selects k targets as the initial cluster center, then calculates the distance from each target to the initial center, and classifies each target to the nearest center point. The K-means algorithm needs to initialize the center, and the number of iterations depends on the input image. The SNIC algorithm does not require iteration, requires only a small amount of memory, and is faster than the SLIC algorithm and the LSC algorithm.

Although the above algorithms have good segmentation results for natural images, some objects are missed when segmenting PCBA images.

SUMMARY OF THE INVENTION

In view of the above deficiencies in the prior art, the actual problem to be solved by the present invention is: how to avoid missing targets when segmenting the PCBA image.

In order to solve the above-mentioned technical problems, the present invention adopts the following technical solutions:

A PCBA-based superpixel image segmentation method, comprising:

S1. Obtain the original image;

S2, enhancing the contrast of the original image to obtain an input image;

S3, quantize the RGB color channels of the input image into m different colors, and select the n colors with the highest frequency in the input image as high-frequency colors to represent all input colors, calculate the significant value of the input image and generate a saliency map;

S4, performing superpixel segmentation on the saliency map, and obtaining superpixel labels;

S5, create a superpixel image based on the superpixel label, and the pixel value of each pixel in the superpixel image is the average gray value of the corresponding superpixel block;

S6, using the maximum threshold segmentation method for the superpixel image to obtain the optimal segmentation threshold, and dividing the original image into a target area and a background area based on the optimal segmentation threshold, and generating a segmentation consisting of the target area and the background area. picture.

Preferably, the calculation formula of the significant value in step S3 is as follows:

P_j为输入颜色c的第j个最近邻颜色的概率，D(c_i,c_j)输入颜色c的第i个最近邻颜色和第j个最近邻颜色在LAB颜色空间内的距离，D为输入颜色c与其k个最近邻颜色的距离之和，

In the formula, S(c) is the saliency value of the input color c, c _i is the ith nearest neighbor color of the input color c, k is the number of nearest neighbor colors, and S(ci ) is the initial saliency corresponding to c _{i .} value,

P _j is the probability of the jth nearest neighbor color of the input color c, D( _{ci ,c j )} The distance between the ith nearest neighbor color and the jth nearest neighbor color of the input color c in the LAB color space, D is the sum of the distances of the input color c and its k nearest neighbors,

Preferably, in step S4, the SNIC superpixel segmentation algorithm is used to perform superpixel segmentation on the saliency map.

Preferably, step S6 includes:

S601, calculate the initial segmentation threshold, the initial segmentation threshold is equal to the mean value of the maximum value and the minimum value of the gray value of the superpixel image;

S602, based on the initial segmentation threshold, the superpixel image is divided into a superpixel target area and a superpixel background area, and the initial segmentation threshold is iteratively updated by an information entropy formula to obtain an optimal segmentation threshold;

S603, dividing the original image into a target area and a background area based on the optimal segmentation threshold;

S604. Generate a segmentation map composed of the target area and the background area.

Preferably, it also includes:

S7. Filter the area where the target area in the segmentation map is smaller than the filtering threshold as the background area.

In summary, the present invention discloses a PCBA-based superpixel image segmentation method, including: S1, obtaining an original image; S2, improving the contrast of the original image to obtain an input image; S3, converting the RGB of the input image The color channels are quantized into m different colors, and the n colors with the highest frequency in the input image are selected as high-frequency colors to represent all the input colors, the saliency value of the input image is calculated, and a saliency map is generated; S4. The saliency map is divided into superpixels, and superpixel labels are obtained; S5, a superpixel image is created based on the superpixel labels, and the pixel value of each pixel in the superpixel image is the average gray value of the corresponding superpixel block; S6, The maximum threshold segmentation method is used for the superpixel image to obtain an optimal segmentation threshold, and based on the optimal segmentation threshold, the original image is segmented into a target area and a background area, and a segmentation map composed of the target area and the background area is generated.

Compared with the existing superpixel segmentation algorithm, the present invention has the following beneficial effects:

The present invention first proposes a superpixel label image method to improve the efficiency of image segmentation. At the same time, the visual system is sensitive to the contrast of visual signals, and the extraction of salient features based on the histogram contrast is more effective, and the SNIC superpixel segmentation algorithm uses color and space information to perform initial image segmentation, although it is a clustering method, but It does not need iteration, and the speed is better than the general clustering algorithm, so it is proposed to use the histogram contrast to extract salient features, and then perform superpixel segmentation. Combined with the superpixel label image generation method of the present invention, the information entropy of the new image is extracted for threshold segmentation. , and finally fine-tune the target and background regions to obtain the final segmentation map.

Description of drawings

1 is a flowchart of a specific implementation of a PCBA-based superpixel image segmentation method disclosed in the present invention;

Fig. 2 is the salient figure example in the present invention;

3 is an example of a superpixel segmentation diagram in the present invention;

4 is a schematic diagram of a superpixel image generation process;

Fig. 5 is a segmentation result diagram of different optimal segmentation thresholds;

Fig. 6 is the final segmentation result graph after filtering target area area is less than 70;

Figure 7 shows the comparison results of saliency maps of different algorithms, row (a) is the original image, row (b) is Ground Truth, (c)-(h) are Watershed, SNIC, Mean-Shift, LSC, SLIC and K- The result graph of means algorithm, (i) is the result graph of the present invention;

Fig. 8 is the PR curve comparison diagram of different algorithms;

Figure 9 is a comparison chart of ROC curves of different algorithms.

Detailed ways

The present invention will be further described in detail below in conjunction with the accompanying drawings.

As shown in Figure 1, the present invention discloses a method for superpixel image segmentation based on PCBA, including:

S1. Obtain the original image;

Because the initial PCBA picture is taken by an industrial camera, the average pixel is about 2777*2138. In order to facilitate subsequent calculations, the original image can be cropped, for example, it can be cropped to a common image processing size of 512*512.

S2, enhancing the contrast of the original image to obtain an input image;

In order to facilitate subsequent feature extraction, image preprocessing can be performed on the original image first, and the stretchlim function in the Matlab software can be directly used to adaptively find a segmentation threshold vector to change the contrast of an image, and a grayscale transformation can be obtained. Adaptive optimal threshold.

S3, quantize the RGB color channels of the input image into m different colors, and select the n colors with the highest frequency in the input image as high-frequency colors to represent all input colors, calculate the significant value of the input image and generate a saliency map;

Specifically, 12 different colors can be selected, and 85 high-frequency colors can be selected to represent all input colors. The visual system is more sensitive to the contrast of visual signals, therefore, it is more effective to extract salient features based on histogram contrast.

S4, performing superpixel segmentation on the saliency map, and obtaining superpixel labels;

S5, create a superpixel image based on the superpixel label, and the pixel value of each pixel in the superpixel image is the average gray value of the corresponding superpixel block;

As shown in Figure 4, for the sake of simplicity of calculation, a method for generating a new superpixel image using label information is proposed in the present invention, wherein each pixel value of the superpixel image is the average gray value of the superpixel block, The calculation formula is as follows:

I(i)=f(O(i)), i=1,2,...,L

Among them, f is the mean function, X and Y are the coordinate set of the label label=i, g(x, y) is the pixel value of (x, y), and O(i) is the gray of the pixel block corresponding to the ith label The degree value, that is, the sum of the gray values of all pixels belonging to the ith label. Arrange these new pixel values into a 28*28 matrix from the original position left to right to form a new image.

In the prior art, the original image is generally operated directly, and the image is segmented by obtaining the information entropy of the original image, or the SNIC algorithm is directly used for direct segmentation. The pixel size of the pixel image is small, but it retains important information such as the edge of the original image, and the calculation speed is greatly improved. Using this method to calculate the information entropy is faster, and the impact of information loss is small.

S6, using the maximum threshold segmentation method for the superpixel image to obtain the optimal segmentation threshold, and dividing the original image into a target area and a background area based on the optimal segmentation threshold, and generating a segmentation consisting of the target area and the background area. picture.

Compared with the prior art, in view of the characteristics that the color information of the PCBA is simple and does not contain complex background information, the present invention uses the superpixel label to generate a superpixel image before the final segmentation, so that the edges of the original image are preserved. and other important information, it will not cause the omission of the target, and reduce the pixels of the image to be processed, thereby reducing the computational cost and improving the efficiency of segmentation.

As shown in Figure 2, when implementing, the calculation formula of the significant value in step S3 is as follows:

P_j为输入颜色c的第j个最近邻颜色c_j的概率，即灰度值为第j个最近邻颜色c_j对应的灰度值的像素点数量占输入图像的像素点总数量的比重，D(c_i,c_j)输入颜色c的第i个最近邻颜色和第j个最近邻颜色在LAB颜色空间内的距离，D为输入颜色c与其k个最近邻颜色的距离之和，

In the formula, S(c) is the saliency value of the input color c, c _i is the ith nearest neighbor color of the input color c, k is the number of nearest neighbor colors, and S(ci ) is the initial saliency corresponding to c _{i .} value,

P _j is the probability of the j-th nearest neighbor color c _j of the input color c, that is, the ratio of the number of pixels with the gray value of the gray value corresponding to the j-th nearest neighbor color c _j to the total number of pixels in the input image , D(c _i , c _j ) the distance between the ith nearest neighbor color and the jth nearest neighbor color of the input color c in the LAB color space, D is the sum of the distances between the input color c and its k nearest neighbor colors,

The RGB color channel can be quantized to obtain 12 different values, and the high-frequency color is selected, and the selected color covers more than 95% of the pixels, and the number of high-frequency colors can be taken as 85.

The color coverage of more than 95% means that the selected color pixel value accounts for more than 95% of the color pixel value of the entire image, and selecting the color to replace the entire image can reduce the amount of calculation.

As shown in FIG. 3 , during specific implementation, in step S4 , the SNIC superpixel segmentation algorithm is used to perform superpixel segmentation on the saliency map.

For the superpixel segmentation of the saliency map, the SLIC algorithm and the SNIC algorithm can be used. The SNIC algorithm is used to perform superpixel segmentation on the acquired saliency map. The SNIC algorithm recalculates each superpixel every time a new pixel is added to the superpixel block. The average value of the pixel block, and the center value of the pixel block is replaced, no iteration is required, and the speed is better than the general clustering algorithm. In addition, the search range of the SNIC algorithm is 4 or 8 neighborhoods of each center point, which can avoid the phenomenon that some pixels are mistakenly divided into the superpixel block due to the short distance from the center point of the superpixel block. Improved segmentation accuracy.

The superpixel distance calculation formula is as follows:

为平衡因子，N为输入图像像素数量，L为超像素个数，M为平衡因子。具体可取平衡因子M＝20，超像素个数L＝800。where d _lab is the color distance in the LAB color space, d _xy is the color distance between two pixels, l, a, and b are the three color feature components in the LAB color space, respectively, and x and y are the coordinate position components,

is the balance factor, N is the number of pixels in the input image, L is the number of superpixels, and M is the balance factor. Specifically, the balance factor M=20, and the number of superpixels L=800.

As shown in FIG. 5 , during specific implementation, step S6 includes:

S601, calculate the initial segmentation threshold, the initial segmentation threshold is equal to the mean value of the maximum value and the minimum value of the gray value of the superpixel image;

S602, based on the initial segmentation threshold, the superpixel image is divided into a superpixel target area and a superpixel background area, and the initial segmentation threshold is iteratively updated by an information entropy formula to obtain an optimal segmentation threshold;

S603, dividing the original image into a target area and a background area based on the optimal segmentation threshold;

S604. Generate a segmentation map composed of the target area and the background area.

The information entropy calculation formula is as follows:

和

分别为目标区域和背景区域的总概率和，且P_b＝1-P_o，p(i,j)为直方图中像素点(i,j)的概率。依据获得的最佳分割阈值对原始图像进行阈值分割，得到分割图。Among them, H _o and H _b are the information entropy of the target area and the background area, respectively, H _I is the information entropy of the picture, when H _I reaches the maximum value after iteration, the optimal segmentation threshold is obtained, (q, r) is the best the coordinates of the segmentation threshold T _max ,

and

are the total probability sum of the target area and the background area, respectively, and P _b =1-P _o , p(i,j) is the probability of the pixel point (i,j) in the histogram. Perform threshold segmentation on the original image according to the obtained optimal segmentation threshold to obtain a segmentation map.

As shown in Figure 6, when implemented, it also includes:

S7. Filter the area where the target area in the segmentation map is smaller than the filtering threshold as the background area.

For the effect of the PCBA-based superpixel image segmentation method disclosed in the present invention, the following performance reference indicators can be used to evaluate:

Among them, S represents the saliency estimation result normalized to [0, 1], G represents the true saliency calibration, W and H are the width and height of the corresponding image, the smaller the MAE value, the better the algorithm performance. TP represents the number of pixels in the target area in the binary saliency result that are consistent with the true value saliency calibration, TN represents the number of pixels in the background area in the binary saliency result that are consistent with the true value saliency calibration, and FP represents the binary saliency The number of pixels that are wrongly classified as the target in the saliency result, FN represents the number of pixels that are wrongly classified as the background in the binary saliency result, and Precision is the precision rate, which refers to all foreground pixels generated by the algorithm. The ratio of correctly calibrated pixels, Recall is the recall ratio, which refers to the ratio of the pixels correctly calibrated by the algorithm in the foreground pixels calibrated by the actual true value. The recall is the horizontal axis and the precision is the vertical axis. The PR curve is obtained. The closer the curve is to the upper right corner, the better the algorithm performance. β ² =0.3, F _β (F-measure) takes both precision and recall into account, and the algorithm performance is proportional to the value of F _β . FPR is False Positive Rate, on the horizontal axis, TPR is True Positive Rate, on the vertical axis, the closer the drawn ROC curve is to the upper left corner, the better the algorithm performance is. The AUC indicator (the Area Under the receiver operating characteristic Curve), that is, the area under the ROC curve, the larger the value, the better the algorithm performance.

Taking the original image size as 512*512, quantizing the RGB color channels to obtain 12 different values, the number of high-frequency colors is 85, M=20, L=800 as an example, the number of labels is 784, and the size of the superpixel image is 28* 28. Table 1 shows the comparison of MAE value and running time under different filtering thresholds.

Table 1 Comparison of evaluation indicators under different filtering thresholds

It can be seen from Table 1 that when the filtering threshold is 70, the value of MAE tends to be stable, and the running time also tends to be stable, so 70 is selected as the filtering threshold of the present invention.

After the image segmentation is completed, the statistical significance map, MAE, F-measure, PR curve, ROC curve and running time in the experiment are obtained, which are used as effective model performance reference indicators. Table 2 shows the MAE, F-measure and Comparison of running time.

Table 2 Quantitative comparison table under different algorithms

It can be seen from Table 2 that the present invention has the highest F-measure value and the lowest MAE value. Although the running time is lower than that of the Watershed algorithm, other indicators are far better than the Watershed algorithm.

As shown in Figures 7 to 9, it can be seen from the above experiments that there are some under-segmentation phenomena in the traditional image superpixel segmentation algorithm when segmenting PCBA images, a superpixel image segmentation algorithm based on PCBA is proposed, which takes into account the Algorithm efficiency and accuracy. Use the gradient histogram contrast algorithm to extract the saliency features, and then use the superpixel segmentation algorithm to segment the saliency map, use the superpixel labels obtained by the segmentation algorithm to generate a new image, extract the information entropy of the new image, and use the information entropy threshold to segment the original image. And filter out the areas where the target area is less than 70 in the segmentation map, and get the final segmentation result map. The final F-measure value is 0.7979, and the MAE value is 0.040567, which is better than the other six algorithms.

The above are only the preferred embodiments of the present invention. It should be pointed out that for those skilled in the art, without departing from the technical solution, some deformations and improvements can be made, and the technical solutions of the above-mentioned deformations and improvements should also be regarded as falling into the scope of protection claimed in this application.

Claims (5)

1. a superpixel image segmentation method based on PCBA, is characterized in that, comprises: S1. Obtain the original image; S2, enhancing the contrast of the original image to obtain an input image; S3, quantize the RGB color channels of the input image into m different colors, and select the n colors with the highest frequency in the input image as high-frequency colors to represent all input colors, calculate the significant value of the input image and generate a saliency map; S4, performing superpixel segmentation on the saliency map, and obtaining superpixel labels; S5, create a superpixel image based on the superpixel label, and the pixel value of each pixel in the superpixel image is the average gray value of the corresponding superpixel block; S6, using the maximum threshold segmentation method for the superpixel image to obtain the optimal segmentation threshold, and dividing the original image into a target area and a background area based on the optimal segmentation threshold, and generating a segmentation consisting of the target area and the background area. picture. 2. the superpixel image segmentation method based on PCBA as claimed in claim 1, is characterized in that, in step S3, the calculation formula of significant value is as follows:

In the formula, S(c) is the saliency value of the input color c, c _i is the ith nearest neighbor color of the input color c, k is the number of nearest neighbor colors, and S(ci ) is the initial saliency corresponding to c _{i .} value,

P _j is the probability of the jth nearest neighbor color of the input color c, D( _{ci ,c j )} The distance between the ith nearest neighbor color and the jth nearest neighbor color of the input color c in the LAB color space, D is the sum of the distances of the input color c and its k nearest neighbors,

3. The superpixel image segmentation method based on PCBA as claimed in claim 1, wherein in step S4, superpixel segmentation is performed on the saliency map using SNIC superpixel segmentation algorithm. 4. the superpixel image segmentation method based on PCBA as claimed in claim 1, is characterized in that, step S6 comprises: S601, calculate the initial segmentation threshold, the initial segmentation threshold is equal to the mean value of the maximum value and the minimum value of the gray value of the superpixel image; S602, based on the initial segmentation threshold, the superpixel image is divided into a superpixel target area and a superpixel background area, and the initial segmentation threshold is iteratively updated by an information entropy formula to obtain an optimal segmentation threshold; S603, dividing the original image into a target area and a background area based on the optimal segmentation threshold; S604. Generate a segmentation map composed of the target area and the background area. 5. the superpixel image segmentation method based on PCBA as claimed in claim 1, is characterized in that, also comprises: S7. Filter the area where the target area in the segmentation map is smaller than the filtering threshold as the background area.

Images