Adaptive Edge Detection Algorithm Based on Improved Grey Prediction Model

Existing edge detection algorithms suffer from inefficient edge localization, noise sensitivity, and/or relatively poor automatic detection capability. Contemporary edge detection algorithms can be improved by targeting these problems to help bolster their performance. Grey system theory can be used to resolve the small data and poor information issues in the local information of uncertain systems. An automatic edge detection algorithm was developed in this study based on a grey prediction model to remedy these problems. Noise characteristics in grey images are used to deploy a noise-filtering algorithm based on local features. A mask with twenty-four edge direction information points (345°) was established based on edge line texture features. By compressing the amplitude of the sequence, the randomly oscillated grey prediction sequence can be converted into a smooth, new sequence. The discrete grey model (1,1) (DGM(1,1)) was established based on this new grey prediction sequence to obtain the grey prediction maximum value. A grey prediction image with enhanced edges was obtained by replacing the pixel value in the original image with the maximum grey prediction value. A grey prediction subtraction image with edges separated from non-edge points was also obtained by subtracting the original image from the grey prediction image. The optimal separation threshold in the grey prediction subtraction image can be determined via the global adaptive threshold selection method. The neighborhood search method was then deployed to remove stray points and burrs from the image after the target was separated from the background, creating the final edge image. Experiments were performed on a computer-simulated phantom to find that both the subjective visual effects and objective evaluation criteria are better under the proposed method than several other competitive methods. The proposed edge detection algorithm shows excellent edge detection ability and is highly robust to noise, though the grey prediction model needs further improvement to optimize the run time.

The sub-pixel edge detection algorithm works by targeting sub-pixel points (i.e., points smaller than the image pixels) for enhanced noise-sensitivity and edge positioning capabilities. There are three types of sub-pixel edge detection algorithms currently in use: moment methods, interpolation methods, and fitting methods. Unfortunately, it is difficult to find a suitable fitting model. The massive amounts of calculation these methods require render them unsuitable for real-time applications [13], [14].
Rotation invariance is an important property of the moment. Moment-based methods can be used to effectively achieve edge detection. Types of moment include space moments, gray moments, and Zernike moments. It is usually necessary to select a large template to improve the edge positioning capability. The edges interfere with each other, however, when there are complex edges inside the template. The adaptive selection of the proper moment template is problematic for this reason [15]- [18].
Other techniques have been designed to strike a balance between edge localization ability and noise-sensitivity ability, such as wavelet transform, mathematical morphology, and neural network methods.
An appropriate multi-scale morphological operator must be selected based on mathematical morphology, at which point the edge positioning ability is strong and the amount of calculation is small, but the edges appear to be rough. To improve the stability of the edge-detector, an appropriate filter can be applied as a pre-operation [19]- [21]. Zheng et al. [19] designed an edge refinement method for this purpose which is effective to some extent but not robust to noise. Edge detection algorithms based on wavelet transform or neural network methods show better edge localization ability and noise-sensitivity performance, but their calculation burden is so large that the processing time is too long for practical application. In the automated artificial intelligence context, scenes must be imaged and the images processed in real time; it is difficult to obtain a priori knowledge of the target, so neural networks are not quick enough for effective edge detection [22]- [26].
Grey system theory was introduced in 1982 by Deng [27] in 1982 to solve the problems of uncertain systems with ''small data'' and ''poor information''. After nearly four decades of development, the grey system theory has been successfully applied in various fields such as economics, military, agriculture, geography, and control system [27]- [30]. A computer does not recognize whether a pixel is an edge in a local area when performing edge detection, so when it processes the image as the computer processes information: with ''incomplete global information''. Grey system theory is very well-suited to edge detection for this reason [2]. At present, the edge detection algorithms based on the grey system theory mainly use three types of grey models, namely grey correlation model, grey prediction model, and grey correlation entropy model.
Ma et al. [31] first applied grey relation analysis (GRA) to the edge detection context successfully in a 2003 study, but their technique required manual selection of thresholds and was sensitive to noise. Li et al. [32] proposed a surface image edge detection algorithm based on GRA in 2012, but their technique also requires manual selection of the appropriate separation threshold. Qiao et al. [33] proposed an automatic edge detection method based on GRA and human visual characteristics in 2016; the detection effect is better than traditional edge detection operators, but the edge continuity is poor and it is prone to over-detection.
Zhang et al. [34] established the grey relation entropy theory based on the theory of Shannon entropy in 1996. Zheng et al. [2] proposed an automatic edge detection algorithm based on grey entropy in 2019 which does have excellent edge detection performance, but is prone to double edges and has an overly time-consuming noise filtering algorithm.
He et al. [35] first utilized the grey prediction model for edge detection in 2005; the algorithm can effectively detect the edges of a given image but it is sensitive to noise. Xie et al. [36] proposed a micro-fiber edge detection algorithm based on the traditional grey prediction model (GM (1,1)) and Niblack model in 2016. A strong and weak edge connection algorithm provides good edge continuity, but the method is prone to edge over-detection for complex images and is not robust to salt and pepper noise.
An adaptive edge detection algorithm based on an improved grey prediction model was developed in this study. The structure of this paper is as follows. Section II introduces related work on grey prediction models and grey prediction sequence selection. Section III presents an adaptive edge detection algorithm based on an improved grey prediction model. Section IV reports a series of edge detection experiments we performed on original images and noisy images to test the proposed method, as well as the disadvantages and future research directions related to the proposed method. Section V provides a brief summary and conclusion.

II. RELATED WORK A. IMPROVED GREY PREDICTION MODEL
The grey prediction model is an important branch of grey system theory. Its purpose is to predict data that is irregular but changes monotonically. The GM (1,1) model is the fundamental and most widely used grey prediction model; many prediction models have been derived based on the original GM (1,1) such as the mean GM (1,1) model (EGM), the original difference GM (1,1) model (ODGM), mean difference model (EDGM), discrete GM (1,1) model (DGM), and grey Verhulst model [37].
The original form of the GM (1,1) model is essentially a difference equation [37], [38]. Because the gray values in the 3 × 3 neighborhood window are discrete data and non-homogeneous exponential sequences, the algorithm proposed in this paper uses the DGM (1,1) model [39]. If the gray values of the pixel in the image neighborhood are very close, the development coefficient of the grey prediction model will be very close to zero; this causes the denominator of the time response sequence of the GM (1,1) model to appear to be zero. The gray values of the pixels in the neighborhood of the image are random within 0-255, and likely have a series of oscillating changes [38], but the reduction formula of the DGM (1,1) model is a strictly homogeneous exponential function [39]. Thus, the predicted data cannot be combined with the oscillating characteristics of the original data and DGM (1,1) model application directly results in substantial error. The original data must be improved using a smoothing operator to meet the requirements of the grey prediction model [38]. The grey prediction model uses a small amount of valid data and grey uncertainty data to predict future data through the accumulation of a sequence. A grey prediction model with a sequence of random oscillations is usually established in the form of an exponential with a monotonical increase. However, the simulated data calculated by the exponential expression does not conform to the original random oscillation sequence, resulting in poor prediction accuracy. It is necessary to compress the amplitude (the difference between the maximum value and the minimum value of the sequence) of the original random oscillation sequence to smooth the original sequence and improve the prediction accuracy [40].
Assume the original prediction data sequence is X (0) = x (0) (k) |k = 0, 1, . . . , n , and x (0) (k) ≥ 0, the maximum difference of data in the original prediction data sequence is Df = X max − X min , where, X max = max{x(k)|k = 1, 2, · · · , n} X min = min{x(k)|k = 1, 2, · · · , n} The proposed smoothing operator is: This operator can compress the amplitude of the data sequence, which effectively improves the smoothness of the modeling sequence. Proof: The prediction sequence after the smoothing operation is and the maximum difference of data in the sequence is then substitution into Eq. (2) yields: Therefore, the smoothing operator (Eq. (2)) can compress the amplitude of the data sequence, which effectively smooths the modeling sequence.
Iterative calculation of Eq. (9) yields Let C = x (0) (2) be the initial value of the oscillation sequence prediction model, then For k > 2, when k is odd, the predicted value can be obtained according to Eq. (8) and Eq. (10).
When k is even, the predicted value can be obtained according to Eq. (8) and Eq. (10). (11) is written by combining odd and even cases; this is the improved grey prediction model used in this study.
According to Eq. (11) (Section II(A)), the grey prediction model based on the grey prediction sequence must pass through the initial pixel  50,163,20,40] can be obtained by changing the initial position of S1; the predicted value is F 2 = 109, and F 2 = 109 − 163 = −54 < F 1 . Initial points should be selected on both sides of the grey prediction consequence to minimize any prediction error in the grey prediction model. The grey prediction model generally requires at least 4 data points for modeling. Here, we input 5 data points to the grey prediction model. To suit the 24 edge texture directions, 24 types of grey prediction sequences were designed in the 5 × 5 neighborhood window representing 0 • , 15 • , 30 • , 45 • , . . . , 180 • , 195 • , . . . , 330 • , and 345 • . The mask is shown in Fig. 1. The corresponding prediction sequences are 0 − 345 • . Red pixels are the initial points of the grey prediction sequence, black pixels are the transition points of the grey prediction sequence, and green pixels are the end points of the grey prediction sequence. These sequences represent 24 distinct edge texture directions, which contain more information than single-direction grey model sequences. The prediction results are thus more accurate and the model is robust to low noise.

III. ADAPTIVE EDGE DETECTION ALGORITHM DESIGN A. NOISE FILTERING ALGORITHM
Figures are easily polluted by noise during generation and transmission, so denoising should be performed prior to edge detection. Salt and pepper noise is the most common and has a relatively large impact on image quality [40]- [42]. In this study, denoising filtering pre-processing was performed for salt and pepper noise specifically. Salt and pepper noise is an extreme value in the image which appears directly to the human eye as black or white. In greyscale images, the values of salt and pepper pixels are 0 or 255. A fast denoising algorithm was designed in this study based on this feature. The processing flow diagram of the proposed fast denoising algorithm is shown in Fig. 2.
(1) Sort the 9 pixels in the 3 × 3 neighborhood. If the median value m 1 of the 9 pixels in 3 × 3 neighborhood is not 0 or 255, then there are fewer than 5 noise points in the 3 × 3 neighborhood window.  To test the proposed algorithm, we compared it against other commonly utilized quick algorithms as shown in Fig. 3. Figure 3(a) shows the original image and Fig. 3(b) shows the image with 0.3 salt and pepper noise; Fig. 3(c) shows the image processed by the median filter method with a 3 × 3 template, Fig. 3(d) shows the image processed by the mean filtering method with a 3 × 3 template, Fig. 3(e) shows the image processed by the Discrete Cosine Transform (DCT) filtering method [43], and Fig. 3(f) shows the image filtered by the proposed algorithm.
As shown in Fig. 3, the image becomes blurred when 0.3 salt and pepper noise is added. The average filtering method and DCT filtering method have the worst filtering effect; the image remains blurry and crowded with noise. After the median filtering process, a portion of the image is still blurred, but the noise processing effect is better than average filtering or DCT. The proposed filtering algorithm shows the best denoising effect and most intact original image among the algorithms tested in this study.
Peak Signal to Noise Ratio (PSNR) is an important index for evaluating image filtering effects [2].  Table 1. The proposed algorithm has higher PSNR than the algorithms we tested, especially when processing low-noise images.

B. DATA PREPROCESSING
The relative simulation percentage error (RSPE) calculated by Eq. (12) of the grey prediction sequence is shown below. RSPE is infinite when the gray value x (0) (1) is zero, so this must be carefully avoided. In addition, all grey prediction sequence data should satisfy the ratio check condition. The gray values of the image pixels can be preprocessed to accomplish this: where x (0) (k)is the original value,x (0) (k) is the predicted value, and n is the length of the grey prediction sequence. The specific data transformation method involves performing a positive offset with respect to the data to prevent a 0 value [44]. The offset is p. Logarithmic processing is performed to satisfy extreme ratio conditions as shown in Eq. (13).
The value range of (x (0) (k) is [0,255]. As z(t) = ln(t) is a monotonically increasing function, the grey prediction sequence satisfies the extreme ratio condition: Five data points were used to form a grey prediction sequence in this study, so n = 5 was plugged into Eq. (14) to obtain a p of 61.94. The transformation model is thus:

C. EDGE DETECTION ALGORITHM
The edge is where the gray level of the image changes drastically. When the grey prediction model uses severely fluctuating data, RSPE will be large. When changing from low to high, as shown in Fig. 1 (a), the predicted value results in an overestimation at the turning point of the trend. When changing from high to low, as shown in Fig. 1(b), the predicted value results in an underestimation at the turning point of the trend. However, when the grey prediction model VOLUME 8, 2020 uses data from non-edge areas, the data does not fluctuate greatly, RSPE is small, and the predicted value is stable. In this study, once we selected an appropriate initial grey prediction sequence value and obtained the prediction image via the grey prediction model, we effectively obtained an edge image by subtracting the original image from the prediction image [36], [45].
To prevent predicted value underestimation ( Fig. 1(b)), we used the 24 edge line directions described in Section II (B) to establish 24 grey prediction models to obtain 24 grey prediction values. The texture direction of the 24 edges is (0 − 345 • ), so RSPE is largest when the maximum value is selected, at which point the edges are enhanced. However, for the non-edge area, the 24 predicted values are almost the same as the actual values (i.e., present no such enhancement). Figure 4(a) shows the grey prediction image using 12 prediction sequences(0 − 165 • ), Fig. 4(b) shows the grey prediction image using 12 prediction sequences (180 − 345 • ), and Fig. 4(c) shows the grey prediction image using 24 sequences (0 − 345 • ). The edges of the bears in the three images have mostly white edge lines while the original image has black edge lines, indicating that the edges have been enhanced. Among the three images, some do now show enhanced edges from top to bottom (such as the back of the bear); in Fig. 4(b), the edges are not enhanced from the bottom to the top (such as the abdomen of the bear) though the edges in both directions in Fig. 4(c) are enhanced. In effect, the proposed method results in significant edge enhancement.
The processing region of the algorithm is a 5 × 5 neighborhood window. The grey prediction model cannot predict the two rows above the image, two rows below the image, two columns on the left of the image and two columns on the right of the image, so we copied the pixels of these regions of the image to the original image pixel value.
According to the original image I 0 , the grey prediction image I p can be obtained by the grey prediction model. We obtained an edge direction data matrix Label corresponding to the maximum prediction value. Because the proposed grey prediction model enhances edge points, the value of these edge points in the subtraction image I d is larger than the value of non-edge points. I d is calculated as follows: The optimal separation threshold Th is obtained by iterative method to separate the target from the background [46].
The adaptive threshold selection method used in this study is operated as follows.
(1) The initial value of the separation threshold T 1 = T max +T min 2 is determined where T max is the maximum value of I d and T min is the minimum value of I d .
(2) T 1 divides I d into part A and part B. The average greyscale µ 1 of part A and the average greyscale µ 2 of part B are calculated to obtain a new separation threshold T 2 = µ 1 +µ 2 2 . (3) The values of T 1 and T 2 are compared and the resolution is set to T 0 . If |T 1 − T 2 | < T 0 , T 2 is the optimal separation threshold Th; otherwise, T 1 is replaced with T 2 and Steps (2) and (3) are repeated until the condition |T 1 − T 2 | < T 0 is met. Finally, all pixels are updated and the edge image is obtained according to Eq. (17). If the pixels are determined to be edge points, their values are set to 1 and edge (i, j) is updated to 1 in the edge image; if the pixels are non-edge points, their values are set to 0 and edge (i, j) is updated to 0.
When edge(i, j) = 0, if there are points edge(i + k 1 , j + k 2 ), k 1 = ±1, 0, k 2 = ±1, 0 in the 3 × 3 neighborhood window and Label(i, j) = Label(i + k 1 , j + k 2 ), k 1 , k 2 = ±1, 0, then edge(i, j) = 1 and the broken edges are connected. The gray value of the pixels connected to the edge line is then close to the edges and there is a ''burr'' on the edge processed by the above steps.
The deburring methods used in this study involve setting one pixel as a starting point and searching outwards for edge points in the 4-connected neighborhood. If there are edge points, the search continues outward. When there are less than 3 points found, it is judged as a ''burr'' and the starting point is deleted. In addition, the area pixels with fewer than 25 connections are deleted via 8-connected neighborhood search method. The small edges and noise are removed and the final edge image is obtained.

IV. EXPERIMENTAL RESULTS AND ANALYSIS
We compared the proposed algorithm in this paper with state-of-the-art algorithms to validate its performance and to explore its flaws and potential room for improvement.

A. ACTUAL IMAGE EDGE DETECTION
In order to verify the performance of the proposed algorithm, we compared it to other state-of-the-art algorithms including the improved Canny algorithm proposed by Rong et al. [12], the traditional GM (1,1) edge detection algorithm (threshold th = 5) proposed by Wan et al. [38], the traditional GRA edge detection algorithm (threshold k = 0.1) proposed by Li et al. [32], the morphological method proposed by Zheng et al. [19], and the Zernike moment method proposed by Peng et al. [16]. Fig. 5 shows the edges detected by various methods on the Berkeley Segmentation Data Set (BSDS500) [47], [48]. As shown in Fig. 5, for simple images (such as those in the first and fourth rows), edges were detected successfully and accurately by various methods. For images with relatively complex scenes (such as those in the third, fifth, ninth, and tenth rows), CANNY+RONG, LI+GRA and ZHENG+ Morphology all detected unneeded background areas, WAN+GM(1,1) missed some edges, and PENG+Zenike detected discontinuous edges and background points. These problems are more serious when the scene is complex or when the grey level of the image does not markedly change (such as the images in the second, seventh, and eighth rows). As shown in Fig. 5(h), the proposed algorithm accurately detects most of the edges; edges are rarely over-detected and edge continuity performance is excellent. This indicates that VOLUME 8, 2020 the proposed method has better edge detection ability than the others we tested, though there some double edges do occur (similar to the LI+GRA).
We also compared the F-Measure results of various algorithms and the proposed algorithm as per the precision and recall information from Eq. (17). The edge detection method is more effective when its F-Measure is higher [51]. We chose k = 2 in this study to obtain the results on the on BSDS500 as shown in Fig. 6. As shown in Fig. 6, the F-measure of the proposed method are higher than other algorithms. In conclusion, the ability of edge detection of the proposed algorithm is better than other edge detection algorithm we compared.
Continuity is an important attribute of edges. In the image segmentation, target recognition, and other edge-based image processing contexts, the continuity of edges crucial significance -it directly affects the image segmentation effects and of target recognition results. The continuity of detected edges must be carefully evaluated. When there are many edge points and few continuous edges, the space range that each edge cross is large and edge continuity is strong. A continuity evaluation index based on the average number of pixels contained in the edge line was designed here accordingly: where ε 1 is the number of edge points, ε 2 is the number of edge points detected by the ground truth (Figs. 5(b)), n 1 is the number of edge lines, and n 2 is the number of edge lines detected by the ground truth (Figs. 5(b)). We calculated a continuity evaluation index for the 10 figures in Fig. 5. A continuity evaluation index comparison is shown in Fig. 7. The continuity evaluation indexes of the proposed algorithm were consistently higher than the other algorithms we tested, which suggests that the continuity of the edges detected by the proposed algorithm is better than other compared state-of-the-art algorithms on original images.

B. NOISE IMAGE EDGE DETECTION
We added salt and pepper noise (with density of 0.05) to the digital image used in Section IV(A) and verified the proposed algorithm by comparison to the other algorithms discussed above. The results are shown in Fig. 8. Under the influence of high level (with density of 0.05) noise, as shown in Figs. 7(c)-(f), RONG+Canny, LI+GRA, WAN+GM (1,1), and ZHENG+Morphology were overwhelmed by noise. As shown in Fig. 7(g), PENG+Moment detected most of the edges but many discontinuous edges occurred. The edge detection effects shown in Fig. 7(h) are similar to Fig. 5(h), which indicates that the proposed algorithm is robust to noise.
The F-Measure curve of various edge detection algorithms is shown in Fig. 9. The F-measure values of the proposed algorithm are also consistently higher than the other algorithms we tested. These results suggest that the proposed algorithm outperforms the others for images that are polluted by noise. In addition, as shown in Fig. 10, the F-measure of the noisy images is very close to that of the original image, indicating that the proposed algorithm has strong noise-sensitivity ability.
A continuity evaluation index of various edge detection algorithms was obtained as shown in Fig. 11. The continuity evaluation index of the proposed algorithm is also consistently higher than those of the other algorithms we tested. These results suggest that the continuity of the edges detected by the proposed algorithm is better than other state-of-the-art algorithms on noisy images.

C. FUTURE RESEARCH
There are two edge lines located on the target and the background, respectively. The mask spans across these edge lines when predicting the points on them. As shown in Fig. 10, the proposed algorithm performs accurate detections regardless of whether the starting point is P1 or P2, as it effectively enhances the edges. In other words, the proposed algorithm can detect double edge lines. Edges are usually are located on the target, however, the proposed algorithm still needs reworking to ensure it can detect double edges. The most effective approach to this is to first extract the region of interest (ROI) of the image, then determine whether the detected edge is on the ROI; if it is, keep it, otherwise, delete it. Many rapid ROI extraction methods are currently available [49].   In the future, we plan to integrate ROI extraction into the proposed algorithm.
The threshold selection of the proposed algorithm is global, the edge detection capability for any areas in a given image  with insufficient lighting is relatively poor (Fig. 13, where the light on the left side of the palace is weak) and may result in edge ''leakage''. We also plan to improve the threshold selection of the proposed algorithm in the future, such as by changing the global threshold selection to local threshold selection to reduce the impact of uneven lighting [50].
Running time is an important index for evaluating an algorithm. The algorithm has better performance when it takes less time to complete the same task as other algorithms. Big O notation reflects the time complexity of algorithm. The time complexity of each algorithm and the running time for edge detection on the image shown in Fig. 3 are listed in Table 2.  Table 2, the time complexity of the six algorithms tested in this study is similar. However, the proposed algorithm and WAN + GM (1,1) use the least squares parameter estimation in the grey prediction model, making them somewhat less efficient than the other four algorithms. In addition, a mask with 24 directions was used in this study to accurately obtain the edge direction -this further increases the running time of the proposed algorithm. In the future, the proposed algorithm must be optimized to adapt to online detection (e.g., by optimizing the mask to reduce the number of grey prediction operations, or optimizing the grey prediction model parameter estimation method) to accelerate the grey prediction operations.

As shown in
The continuity of the edges detected by the proposed algorithm is better than the other algorithms tested here, but did result in some missed edges in the test. This problem intensifies when processing images with many weak edges, so the proposed algorithm should to be improved with an edge connecting algorithm. Existing edge connection technologies include morphology-based edge connection algorithms and gradient-based edge connection algorithms. The pixel edge trend can be obtained on several masks with direction information when obtaining the maximum predicted value at any pixel. When there is at least one edge point in the edge trend in the 3 × 3 neighborhood window, the point is determined to be an edge point and set to one.

V. CONCLUSION
This paper proposes a novel image edge detection algorithm based on the grey prediction model. An improved grey prediction model was established based on the original grey prediction model and the gray value of the pixels in the neighborhood window. A grey prediction sequence mask with 0-345 • edge texture direction was then established. The adaptive threshold was selected via global iteration method, then desultory points and burrs were removed via neighborhood search method to complete the automatic adaptive edge detection of grey images.
We performed a series of experiments to test the proposed algorithm's edge detection ability and noise-sensitivity ability. The results show that it outperforms several other similar algorithms and can effectively perform automatic adaptive edge detection. It also showed excellent edge positioning ability and noise-sensitivity ability, which can provide reference for artificial intelligence algorithms such as target recognition and image recognition. However, further research is yet needed to overcome its lack of robustness to uneven illumination and double-edge problems.
ZHEN ZHENG was born in Hubei, China. He is currently pursuing the Ph.D. degree in mechanical engineering with the Nanjing University of Science and Technology (NJUST), Nanjing, China. His current research interests include laser detection, laser target recognition, and image processing.
BINGTING ZHA was born in Jiangxi, China. She received the Ph.D. degree in armament science and technology from the Nanjing University of Science and Technology (NJUST), Nanjing, China, in 2015.
She is currently an Associate Professor with the School of Mechanical Engineering, NJUST. Her current research interests include underwater laser detecting and laser applications.
HAILU YUAN (Member, IEEE) was born in Shanxi, China. She is currently pursuing the Ph.D. degree in armament science and technology with the Nanjing University of Science and Technology (NJUST), Nanjing, China. Her current research interests include underwater laser scanning imaging technology and laser target detection.
YOUSHI XUCHEN was born in Zhejiang, China. He is currently pursuing the Ph.D. degree in armament science and technology with the Nanjing University of Science and Technology (NJUST), Nanjing, China. His current research interests include laser detection and laser anti-dust interference.
YANLIANG GAO was born in Hebei, China. He is currently pursuing the M.S. degree in armament science and technology with the Nanjing University of Science and Technology (NJUST), Nanjing, China. His current research interests include laser detecting and laser applications.
HE ZHANG was born in Henan, China. He received the Ph.D. degree in measurement technology and instruments from the Nanjing University of Aeronautics and Astronautics, Nanjing, China.
He is currently a Professor with the School of Mechanical Engineering, Nanjing University of Science and Technology (NJUST), China. He is the Director of the Institute of Mechanical and Electrical Engineering, NJUST, and the Associate Director of the ZNDY National Defense Key Laboratory. His current research interests include mechatronics and weapon system applications. He is an Editorial Board Member of the Journal of Detection and Its Control.