The purpose of image segmentation is to achieve effective extraction of objects of interest in images. In thermal infrared human target detection, the quality of image segmentation directly affects the execution accuracy and efficiency of subsequent links. Due to the weak nature of thermal infrared images, the information that can be used to segment human objects in images is less and difficult to use, which poses a huge challenge to the performance of image segmentation methods. In this chapter, on the basis of summarizing the commonly used segmentation methods of thermal infrared images, several human object segmentation methods based on active contour model are analyzed and discussed.

 

Thermal Infrared Image Segmentation Method

 

Image segmentation is to classify the pixels in the image into several disjoint and statistically uniform sub-regions (or sets) according to certain characteristics of the image, so as to determine the position of the object of interest in the image. If P is used to represent the pixel set in the image, S;

 

1) Edge detection method

The edge detection method uses edge detection operators, such as Prewitt operator, Sobel operator, Canny operator, etc., to detect the boundary of the image object, so as to segment the image object. Because these operators are often based on image gradient information, they are more effective for images with simple scenes and better imaging quality, but for images with complex scenes, they are prone to problems such as incomplete target boundaries, many false edges, and susceptible to noise.

 

The performance of classical edge detection operators on thermal infrared image segmentation is often poor. For this reason, researchers have proposed several improved operators. Taking the contour saliency map (CSM) operator [58] as an example, in order to deal with the blurred human body edge and halo effect in the thermal infrared image, it estimates the background of the image by suppressing too large and too small The gradient amplitude of the image can weaken the halo at the edge of the human body, thereby reducing the occurrence of false edges, but because the outline of some parts of the human body is quite weak, the problem of boundary fracture is still unavoidable. Compared with edge detection operators based on gradient detection, non-gradient edge detection operators such as the smallest univalve segment assimilating nucleus (SUSAN) operator [59], phase congruency (PC) detection, etc. The detectors can deal with noise and weak edges more effectively, so they are more accurate and reliable for edge detection of human targets in thermal infrared images.