Infrared image segmentation method for power equipment based on improved cluster region growth

For overheating defects of power equipment, the use of infrared technology is widely popular at present, which is less costly and more efficient than the traditional manual detection of thermal defects of power equipment. However, infrared images have the nature of concentrated intensity and low contrast, and picture segmentation has always been a difficult point. This paper proposes a combination of K-mean clustering and improved region growing algorithm, compared with the traditional region growing algorithm, solves the need to manually select the seed point to produce uneven gray scale and over-segmentation and under-segmentation, etc., through the K-mean clustering algorithm to automatically select the number of seeds as well as the seed node, and the introduction of Canny operator to reduce the error in order to achieve a better segmentation effect. Finally compare other algorithms fuzzy C-mean segmentation and fuzzy threshold segmentation.


Introduction
Infrared thermography technology is widely used in the field of power equipment and temperature defect detection due to its fast temperature measurement response, non-contact measurement, wide measurement range, and high temperature resolution.In automatic infrared fault detection, different types of power equipment have different fault characteristics, and typically, the overheated regions of equipment are intertwined with normal regions and the background environment, posing challenges for segmentation.Therefore, image segmentation of power equipment is currently being extensively researched [1].At present, infrared image segmentation methods for power equipment are in the developmental research stage, and thousands of image segmentation methods have been proposed from the research.The more conventional segmentation methods include threshold-based segmentation methods [2], edge detection-based segmentation methods [3] and region-based segmentation methods [4][5].There are also segmentation hair methods based on specific theories.Existing studies use segmentation based on FAsT-Match algorithm, FAsT-Match algorithm in the visible image approximate template matching, and then in the infrared and visible image through the approximate affine transformation to locate the approximate region of the target in the infrared image, and finally segmentation algorithm for the approximate area of the segmentation, the defects of the method is that the accuracy is very low.The infrared segmentation method using C-mean clustering is very popular, and its algorithm has high accuracy in segmentation effect [6], but it is time-consuming and inefficient [7].In response to the challenges of the existing infrared image segmentation algorithms for power equipment, such as the difficulty in determining empirical thresholds and inaccurate segmentation in the presence of complex lighting conditions, a novel infrared image segmentation method for power equipment based on K-means clustering and improved region growing is proposed.By considering the characteristics of power equipment infrared images, the K-means clustering algorithm is employed for automatic seed point selection.Additionally, the Canny operator combined with the gradient operator is used to calculate the gradient magnitude of each pixel, which serves as an additional directional constraint to achieve effective extraction of power equipment targets in the infrared images.Experimental results demonstrate that the proposed method exhibits significant advantages in terms of detailed representation and accurate segmentation of power equipment infrared images.The segmentation Mean Error (ME) value is reduced by an average of 84.76% and 76.19% compared to the fuzzy C-means segmentation method and the fuzzy threshold segmentation method, respectively, indicating a superior segmentation accuracy.

K-means clustering segmentation algorithm selects the seed points
K-means algorithm is a kind of clustering algorithm, different scholars in different fields independently proposed this clustering idea.Although the K-means algorithm has been proposed for half a century now, it is still one of the most widely used clustering algorithms, and it is also applicable to the field of image segmentation.Considering a digital image as a dataset, suppose there are n dimensional vectors and the dataset is X = x 1 , x 2 , x 3 ⋯x n The data set is X be divided into K clusters, forming a subset C = c 1 , c 2 , c 3 ⋯c k The steps of the algorithm are as follows: (1) Randomly select K objects from dataset X as initial cluster centers (seed points); (2) Calculate the distance between each object and these centroid objects based on the mean (central object) of each cluster, and reassign the corresponding objects according to the minimum distance; (3) Compute the mean of each cluster that has changed, and calculate new cluster centers (seed points) based on these means; (4) Run the loop (2) to (3) steps until each cluster center (seed point) is not changing.The traditional region growing method requires manual selection of seed points, which can result in under-segmentation and over-segmentation due to uneven grayscale and fuzzy boundaries.In this paper, by utilizing the K-means clustering segmentation algorithm for automatic seed point selection, this situation can be significantly improved, leading to a reduction in image segmentation errors.

Improved region growth segmentation algorithm
Region growing algorithm is a commonly used basic method in image segmentation.The main idea of the region growing algorithm is to select initial seed points and check whether adjacent pixels have similar properties to the seed points, determining whether to merge the adjacent pixels into the region where the seed points are located.This process continues by checking the neighboring pixels of the merged region pixels until no new pixels are merged into the region.Let the region a contain pixels with mean gray value: The comparative test for whether pixels are merged or not is expressed as: The value should not be too small, too small will lead to under-segmentation due to gray scale changes, too large will lead to over-segmentation due to blurred edges.

Analysis of experimental results
Simulation experiments are conducted on the grayscale maps of infrared images of transformer bushings and static contacts of circuit breakers using MATLAB 2019 under Windows 10 system using the proposed in this paper Infrared image segmentation method for power equipment based on K-mean clustering with improved region growing.To validate the algorithm of this paper compare fuzzy Cmean segmentation and fuzzy threshold segmentation., where B S and Q S are the sets of background and foreground pixels of the binarized standard segmentation map of the power equipment.B F and Q F are the sets of background and foreground pixels of each thresholding segmentation image, and B S ∩ B F and Q S ∩ Q F denote the sets of pixels corresponding to the standard segmentation map and the thresholding segmentation map.ME values are calculated after image processing with the segmentation method of this paper and fuzzy C-mean segmentation and fuzzy threshold segmentation, respectively.The ME value statistics of each calculation method are shown in Table 1.

conclusion
This paper addresses the issue of manual selection of initial seed points in the traditional region growing method and proposes an automatic seed point selection method using the K-means clustering algorithm, taking into account the characteristics of power equipment infrared images.It is also noted that the traditional region growing criteria based on regional grayscale difference can lead to undersegmentation and over-segmentation due to uneven grayscale and fuzzy edges.To overcome this, the Canny operator is used in combination with the gradient operator to calculate the gradient magnitude of each pixel, providing an additional directional constraint.Experimental results demonstrate that the proposed method effectively and accurately segments the temperature field of power equipment infrared images.The infrared segmentation results obtained using this method exhibit significant advantages in terms of detailed representation and precise segmentation.The Mean Error (ME) value of segmentation is reduced by an average of 84.76% and 76.19% compared to the fuzzy C-means segmentation method and the fuzzy threshold segmentation method, respectively, indicating superior segmentation accuracy.

Figure 1 .
Figure 1.Results of different segmentation methods for transformer bushings.

Figure 2 .
Figure 2. Results of different segmentation methods for circuit breaker static contacts.It is difficult to distinguish the performance advantages and disadvantages of each algorithm based on subjective judgment, and the use of the misclassification rate (ME) as an objective index of image segmentation can further compare the segmentation effect of the two algorithms.The expression of ME is: = 1 − ∩ + ∩ +

Table 1 .
Comparison of ME values for each algorithm.After data comparison, it can be seen that in terms of ME value, the method of this paper has obvious improvement comparing fuzzy C-mean segmentation and fuzzy threshold segmentation, and the segmentation effect is obviously superior.The distribution of ME value is reduced by 84.76% and 76.19% comparing fuzzy C-mean segmentation and fuzzy threshold segmentation.