The threshold segmentation method assumes that each target area in the image is composed of many pixels with similar gray values and there is a significant difference in the gray levels between the foreground target and the background or between different foreground targets. The target is separated from the background. Because the gray level of the human target in the thermal infrared image is generally higher than that of the background, threshold segmentation is a simple and effective method, but the human target obtained by threshold segmentation is usually incomplete and requires more post-processing.

 

Threshold segmentation methods can be divided into histogram shape analysis methods, clustering-based methods, entropy methods, target attribute-based methods, spatial information methods, and local adaptive thresholding methods according to the source of information used to determine the threshold. The histogram shape analysis method determines the segmentation threshold by analyzing the peak, valley, curvature, etc. of the smooth histogram; based on the clustering method, the gray level of the image is clustered into two categories: background and foreground (target), or two Gaussian mixture models are used. Fitting determines the segmentation threshold; the entropy method determines the segmentation threshold based on the information entropy of the foreground and background of the image or the image cross-entropy before and after thresholding; the target attribute-based method determines the segmentation threshold based on the similarity between the gray level and the binary image Measures, such as fuzzy shape similarity, edge consistency, etc. determine the segmentation threshold; the spatial domain information method determines the segmentation threshold according to the high-order probability distribution and (or) correlation between pixels; the local adaptive thresholding method determines the local image segmentation according to the local characteristics of the image threshold.

 

Threshold segmentation methods can also be divided into parametric methods and non-parametric methods. The parametric method performs parameter fitting according to the first-order statistical characteristics of the image to determine the image segmentation threshold. The main shortcomings of the parametric method are: ① The optimal threshold is not always located at the intersection of Gaussian distribution; ② If the image histogram itself is unimodal or the two types of distributions have highly overlapping characteristics, their effectiveness will be seriously degraded; ③ The histogram distribution of some types of images in nature cannot be described by Gaussian distribution; ④ If the number of categories in the process of image thresholding increases, the amount of calculation will increase exponentially. The non-parametric method implements threshold segmentation by optimizing some posterior criterion functions, and does not involve the estimation of various distribution parameters of the image in this process. Typical non-parametric methods include p-tile method, Otsu method and entropy method (maximum entropy, Renyi entropy, Tsallis entropy, cross entropy, etc.). Among them, the p-tile method assumes that the image is composed of a black target and a bright background, and assumes that the area of the target area is known, then the optimal threshold is to make the pixels contained in the target area in the segmented image account for at least (100-p)% of the gray level class. The Otsu method determines the threshold by maximizing the variance between classes. For images whose histograms show obvious peak-valley distribution characteristics, the segmentation effect is very good, but for images whose histogram distribution is unimodal or the peak-valley characteristics are not obvious, the segmentation effect is not good. good. The mathematical mechanism and physical meaning of the entropy method are relatively clear, and the adaptability and practical performance are better.

 

For thermal infrared images, researchers have also proposed many special criteria for determining the threshold, such as the difference between the maximum gray value of the image and a preset constant, the last (non-saturated) local minimum in the image histogram, the image mean and The weighted value of the maximum gray value, the value determined according to the ratio of the area under the smooth part of the curve in the histogram to the total area under the curve, the value determined according to the fuzzy Havrda-Charvát entropy, the value determined according to the image mean and standard deviation, etc.