Infrared image edge extraction based on human eye micro-motion mechanism

The edge of an image is the intersection between regions with different attributes of the image, and it is also the place where the regional attributes suddenly change. The edge of an image contains most of the information of the image, and is an important basis for image recognition and classification, as well as an important part of image processing research. Scholars from various countries have proposed many image edge extraction algorithms, such as the classic Sobel, Prewitt, LOG, Canny and other algorithms [1]. The principles of these algorithms are simple and easy to implement, but since most of them are based on differential gradient operators of local windows, they are sensitive to noise and are therefore not suitable for processing images that are severely disturbed by noise. With the development of computer technology and the application of other disciplines in image processing, scholars have proposed many edge extraction algorithms based on new theories, such as edge extraction algorithms based on morphology, fuzzy mathematics, wavelet transform, and fractal [2-4]. Each algorithm has its own characteristics and a certain scope of application. Due to the particularity of the infrared imaging mechanism, infrared images generally have the characteristics of low contrast, blurred edges, and low image signal-to-noise ratio, making it difficult to extract the edges of infrared images. Since the micro-motion of the human eye has the characteristics of super-resolution, it provides a good inspiration for infrared image processing. This paper studies the extraction of infrared image edges based on the mechanism of human eye micro-motion and verifies it through experiments. The experimental results show that the infrared image edge extraction algorithm based on the human eye micromotion mechanism can not only accurately extract the edge of the image, but also effectively reduce the pseudo-edges in the image, and has a good edge extraction effect.
1 Principle of human eye micromotion imaging
As early as 1952, DITCHBURN and GINSBORG et al. noticed that the human eye has unconscious tiny movements in the fixed state, namely human eye micromotion [5,6], which can be divided into three modes: high-frequency vibration, drifting movement and flicker. Since the amplitude of the first two is not large and the current eye detection technology is difficult to measure accurately, the human eye micromotion mentioned here mainly refers to flicker. Foreign scholars have conducted a lot of experimental research on eye micromotion. The experiment shows that when the human eye stops all eye movements, the static image will suddenly become blurred and then disappear. There is a direct relationship between eye micromotion and the generation of vision, and binocular micromotion is better than monocular micromotion. In recent years, Condem et al. [7] have further revealed that eye micromotion is directly related to image disappearance by measuring the probability, rate and amplitude of micromotion before and after image disappearance or reappearance, and that micromotion is causally related to the clarity of fixed vision images.
The imaging principle of human eye micromotion [8,9] is as follows: human eye micromotion makes the information of interest fall more accurately into the fovea of ​​the retina. The more accurately the information falls into the fovea, the stronger the image sharpness perceived by the human eye. Human eye micromotion can enhance the reaction of the near fovea, and they can correct the information of the fovea through the lateral connection of the inhibitory neural network. If the correction is not ideal, human eye micromotion can make the information return to the fovea more accurately, and adjust the micromotion amplitude according to the size of the information scale, and repeat this process until the information is accurately identified. The smaller the information scale, the smaller the micromotion amplitude, and vice versa. At the same time, as the micromotion rate increases, the sharpness of the corresponding retinal field of vision will be stronger. This paper approximates the micromotion mechanism of the human eye by translating the image and applies it to the edge extraction of infrared images.
2 Basic process of the algorithm
2.1 Basic principle of the algorithm
The human eye mainly relies on the micro-motion mechanism to distinguish the edge of the image. The smaller the amplitude of the eyeball micro-motion, the finer the edge of the image. As the amplitude increases, the edge line of the image becomes thicker, but the ability to highlight large-scale edges is stronger than that of small amplitudes. The theory of human eye micro-motion has a strong edge extraction ability. The basic principle of the algorithm is shown in Figure 1 [10]. Combined with the sensitivity of retinal ganglion cells to direction, the direction of human eye micro-motion is first selected, and then the edge image of the micro-motion direction is calculated through the micro-motion image. After that, the edge images of each micro-motion direction enter the competition stage. The result of the competition is the optimal overall grayscale edge image of each micro-motion direction. Finally, binarization is performed to generate a binary edge image.
Suppose the original image is f(x,y), then the micro-motion edge image of the image in a certain micro-motion direction is expressed by the following formula:
g(x,y)={f(x,y)-f?兹(x+k1·?住x,y+k2·?住y)} (1)
In the formula, ?住x, ?住y represent the distance unit of the image moving in the x and y directions respectively, k1, k2 represent the size of the movement, ?兹 represents the direction of the movement, and the calculation formula is:
?兹=arctan(k2/k1) (2)
Let r(x,y) represent the output of the competitive network, then the competitive grayscale edge image can be expressed as:
r(x,y)=max{gi(x,y)|?坌i} (3)
Set the threshold t to obtain the binary edge image:
R(x,y)=1 r(x,y)≥t0 r(x,y) 2.2 Calculation step
(1) Input the original image A, and generate the comprehensive micro-motion image F by translating the original image A in the micro-motion direction. F = [Ah, Av, Ad], where h, v, d represent the horizontal, vertical and oblique directions respectively. In this paper, image A is translated in eight directions, and the moving distance is one pixel.

(2) Calculate the edge image H of each micro-motion direction:
Ci=Fi-A, i=h,v,d (5)
(3) Calculate the competitive grayscale edge image H:
H=max(Ci), i=h,v,d (6)
(4) Requantize the competitive grayscale edge image H to [0, 255].
(5) In order to reduce the generation of pseudo edges, perform mean filtering on the competitive edge image H:
G=mean(H) (7)
(6) For the grayscale edge image after quantization filtering, non-maximum suppression and double threshold detection edge connection processing are used to obtain a binary edge image.
2.3 Non-maximum suppression
Directly binarizing the competitive grayscale edge image after quantization filtering cannot accurately extract the edge of the image. Therefore, it is necessary to perform non-maximum suppression on the amplitude of the competitive grayscale edge image after quantization filtering to further determine the edge points. If the edge intensity G(i,j) of the pixel (i,j) on the image G(x,y) is less than the edge intensity of the two adjacent pixels along the translation line, the pixel is considered to be a non-edge point and its grayscale value is set to 0. That is, the point with the largest local amplitude change is retained, and the ridge band in the amplitude image is refined.
2.4 Double threshold detection and edge connection
Since both noise and edges in the image belong to high-frequency parts, a large part of the edge image after non-maximum suppression processing is still pseudo-edge points belonging to noise, so denoising must be performed [7]. This paper adopts a high and low double threshold method to implement this denoising process. Set high and low thresholds. The edge image after high threshold processing can remove most of the noise and obtain a clear edge with a larger size, but at the same time loses some useful edge details; the image after low threshold denoising retains more information and can retain subtle edges, but produces more pseudo edges. After double threshold processing, two binary edge images with different features can be obtained. Based on the high threshold edge image and supplemented by the low threshold edge image, edge connection is performed to achieve the final image edge extraction.
3 Experimental results
The experimental results are shown in Figure 2, where Figure 2(a) is an original infrared image of size 405×401. The overall grayscale value of the image is low, it is seriously affected by noise, the contrast is not strong, and the edges are blurred; Figures 2(b) and 2(c) are the edge extraction effects of Sobel and morphological algorithms respectively, and Figure 2(d) is the edge extraction effect of the algorithm in this paper.
It can be seen from Figure 2 that the edge extraction effect of the Sobel algorithm is poor, the target edge fracture phenomenon is more serious, and it is more sensitive to noise; although the morphological edge extraction extracts the edge of the target more completely, it produces a large number of pseudo edges; the edge extraction effect of the target edge is more ideal when the algorithm in this paper is used to extract the edge, and it is not sensitive to noise, and the extracted pseudo edges are less (the pseudo edges around the image are generated during the translation process of the image, which does not affect the subsequent processing of the image). Infrared
images are seriously polluted by noise and have blurred edges. It is difficult to extract edges using traditional edge extraction algorithms. Based on the basic principles of human eye micro-motion visual imaging and combined with the characteristics of infrared images, this paper studies edge extraction. Experimental results show that the edge extraction algorithm based on the human eye micro-motion visual imaging mechanism can quickly and accurately extract the edges of infrared images, and can better suppress the generation of pseudo edges, achieving good edge extraction effects.
References
[1] Zhang Xiang, Liu Meijie, Chen Liwei. Edge extraction method based on mathematical morphology [J]. Journal of University of Electronic Science and Technology of China, 2002, 31(5): 492-495.
[2] Shi Zhengang, Tao Yaodong, Gao Liqun. A fuzzy enhanced image edge extraction algorithm [J]. Microcomputer Systems, 2008, 29(11): 2145-2148.
[3] Pu Xiaoqin. Image recognition based on multifractals [D]. Xi'an: Northwest University, 2009.
[4] DITCHBUM R W. GINSBORG B L. Vision with a stabilized retinal image [J]. Nature，1952, 170(4314): 36-37.
[5] DOWNTON A, CROOKES D. Parallel architectures for image processing [J]. Electronics & Communication Engineering Journal, 1998, 10(3): 25-26.
[6] MARTINEZ C, STEPHEN L M. Microsaccades counteract visual fading during fixation[J]. Neuron, 2006(49):297-305．
[7] Yan Jinming, Li Yanjun, Zhang Ke. Infrared image edge extraction based on human eye visual mechanism[J]. Firepower Command and Control, 2009, 34(1):60-62.
[8] Li Zhijun, He Ming. Real-time infrared image edge extraction based on human eye micromotion mechanism[J]. Optoelectronic Engineering, 2008, 35(12):117-121.
[9] Wang Li. Research on visual mechanism and its application in infrared imaging guidance[D]. Xi'an: Northwestern Polytechnical University, 2003．
[10] GONZALES RC, WOODS R E. Digital Image Processing[M]. Beijing: Electronic Industry Press, 2005．