An improved local binary pattern based edge detection algorithm for noisy images

Abstract

Local Binary Pattern (LBP) is considered as an effective image descriptor as it is based on joint distribution of gray level differences. The main attributes of LBP are discriminatory power, robustness to brilliance change, simplicity and computational efficiency. In contrary LBP is highly sensitive to noise, rotation, non-rigid deformation, view point variations and scaling. Therefore, in the present work an improved version of LBP i.e. ILBP is proposed to overcome the limitations of basic LBP. ILBP replaces the fixed-weighted matrix of basic LBP by a pixel difference matrix. The proposed method is assessed on synthetic as well as real-time images. The results obtained are compared with LBP and other state-of-the-art edge detection techniques like HLBP, Canny and Sobel methods. The results reveal that performance of ILBP is superior to other edge detection methods under consideration. Further the proposed technique is highly efficient for noisy, blurred and low pixel valued images.

Keywords

Digital radiography imaging edge extraction LBP Canny Sobel

1 Introduction

An edge is formed in an image if there is a change in color, shade, texture and light. These edges are used to determine orientation, surface properties, size and depth of an image. Thus edge detection is an essential part of image processing and computer vision system. The process of edge detection involves filtering of irrelevant information from the image so as to select an edge point. It is employed to simplify the analysis of images by minimizing the amount of data to be processed. Major application areas of edge detectors include medical imaging, intelligent traffic system, robotics, military, geography and remote sensing [1 –5] etc. Various edge detection techniques are proposed in literature e.g. Canny [6], Sobel [7], Roberts [8], Kirsch [9] etc. which are computationally efficient and provide effective results for good quality images.

However, in the presence of noise and artifacts, they provide poor efficiency with a possibility of producing broken edges [10]. Further they often fail to identify boundaries of complex objects having sophisticated textures like remote sensing and medical images.

Recently various edge detection schemes are developed to overcome the shortcomings of classical edge detection methods. Mendoza et al. [11] introduced a novel hybrid method of edge detection based on fuzzy interface scheme and Sobel filter. In this method fuzzy rules are designed to estimate the normalized values of edges. The authors [12] also designed hybrid technique using fuzzy logic, neural network and Sugeno integral for face recognition. Melin et al. [13] presented a morphological gradient and fuzzy logic based technique for identifying edges in digital images. The technique provides satisfactory results by preserving more details than the conventional techniques. Wang et al. [14] claimed another innovative edge detection technique using fiber bundle geometry approach and Lie-algebra valued gauge field. The performance of suggested approach is better than Canny and Hui method. Chongke et al. [15] developed edge detection technique for art image using a dynamic mode decomposition algorithm and colour space method to remove shadow from an image. Deng et al. [16] introduced an edge detection technique using fusion technology which combines improved Sobel operator and wavelet transformation for noisy images. Simulation results show that fusion technique provides complete details and accuracy of the edge detection is improved. Murala et al. [17] suggested a local mesh peak-valley edge pattern (LMePVEP) for MRI and CT image indexing and retrieval. LMePVEP uses forward and backward first-order derivatives for extracting the relationship between middle pixel and neighbour pixels. Subrahmanyam et al. [18] presented content based image retrieval (CBIR) for object tracking applications. CBIR is performed by conducting four sets of experiments, three of which are utilized for image retrieval and fourth for object tracking. Meng et al. [19] designed automatic edge detection techniques based on local adaptive Canny and modified circular Hough transform to detect particle size distribution. Simulation results reveal that suggested techniques are robust, reliable and detect the particles with high efficiency. Lin et al. [20] presented quasi high pass filter based on local statistics for edge detection of medical image. The technique provided thin and unbroken edges of CT scan, X-ray and MRI images. Biswas et al. [21] proposed modified Moore-Neighbor method for feature extraction and tracing boundary of noisy images. The performance of designed method is better as compared to Moore-neighbor, improved Sobel detector and active Canny techniques for noisy images. Guada et al. [22] developed an image divide and link algorithm using hierarchical graph partition approach for edge detection. The experimental results prove the superiority of proposed technique as compared to Canny and Sobel methods for BSDS 500 data set. Chaira [23] suggested triangular operators based fuzzy morphological approach for edge detection of normal, noisy and gradient images. The devised technique provides better results as compared to Xu’s method, Lukasiewicz operator, Hamacher and product operators. B. Bi et al. [24] formulated an effective edge extraction method for blurry radiography image using local binary pattern improved by embedding H function and counting scheme. Gaidhane et al. [25] designed a simple and efficient edge detection technique based on hyper smoothing function and local binary pattern to detect defects in metal sheets.

Literature survey reveals that any spatial edge detection technique is not capable of identifying meaningful edges on all scales. Such techniques require large computation for multiscale images. Further discrete gradients create an imbalance in the classical edge detectors [26]. It is very important to find edges without breaking the edge stability, which necessitates the requirement of a good technology. The correct information of edges is another requirement for successful edge detection system. Therefore this article focuses on developing an algorithm using an efficient counting scheme and local binary pattern for edge detection in digital noisy images and compressed images. The aim is to achieve thin continuous edges effectively with minimum jagged edges. The performance of proposed method is tested on compressed images, image corrupted with salt and pepper noise of different density and blurred images. LBP, HLBP, Canny and Sobel edge detection techniques are also designed and tested for comparative study. The rest of the paper is organized as follows. Section 2 reviews the background literature of orthogonal local binary pattern. Section 3 presents the improved LBP technique. Section 4 presents the experimental results of algorithms implemented on compressed and noisy images. Finally, Section 5 concludes the work.

2 Local binary pattern

The local binary pattern method is introduced for texture classification by T. Ojala [25]. This technique is successfully employed in various fields of image processing like text recognition [27], face recognition [28], image forgery [29] and image area commentators [30, 31]. LBP gained attraction primarily due to its low computation complexity and high robustness for local adaptation. Basic LBP computes a gray value corresponding to each pixel by comparing the middle pixel with its surrounding pixels. The surrounding pixels whose gray value is found greater than and equal to the middle pixel are encoded as ‘1’ else replaced by ‘0’ and a binary code is obtained. The received binary pattern is further converted into a decimal code by multiplying it with a predefined weight matrix, W_m. The weight matrix is obtained by filling each matrix element in a clockwise fashion with progressive powers of two except the centric pixel. Thus jth element possesses a value corresponding to 2^j where j = 0, 1, 2 … 7.

The LBP value of current pixel is calculated as:-

${LBP}_{j} = \sum_{j = 0}^{m - 1} A (q_{j} - q_{c}) 2^{j}, where j = 0, 1, 2, \dots, m - 1$ (1)

and $A (q_{j} - q_{c}) = {\begin{matrix} 1 & {if q}_{j} \geq q_{c} \\ 0 & otherwise \end{matrix}$ (2)

where q_c and q_j are the values of middle and surrounding pixel respectively. The implementation process of basic LBP for one pixel is illustrated in Fig. 1. The method is repeated for all pixels to obtain the processed image. Basic LBP offers several attributes like ease of implementation, economic computation and efficient classification. In contrary it is quite sensitive to noise and not able to depict the degree of difference while calculating the LBP values [25]. As an illustration, basic LPB may provide same LBP values for two different pixel matrices (Fig. 2). In this context, various researchers proposed enhanced versions of binary LBP method like local trinary pattern (LTP) [32], local quinary patterns (LQP) [33], pyramid local binary pattern (PLBP) [34], dominant local binary pattern (DLBP) [35], LBP variance (LBPV) [36] and local multiple patterns (LMP) [37]. These methods use multiple threshold values for comparison purpose which results in increased accuracy at the cost of large computational effort. Further a suitable optimization technique is also required for selection of optimal thresholds. Consequently, all these variants make the task of image processing more cumbersome. Therefore in the present work efforts are made to improve the basic LBP method while retaining its main attributes.

Fig.1

LBP and ILBP operation.

Fig.2

Example of different regions with LBP and ILBP.

3 Improved local binary pattern (ILBP)

The sensitivity to noise and degree of difference problem may be solved with the help of improved LBP technique. Sensitivity to noise is reduced by recording T_v of the neighboring pixels, whose gray values are equal to or greater than the center one. $T_{v} = \sum_{j = 0}^{7} K (q_{j} - q_{c}), K (λ) = {\begin{matrix} 1 & λ \geq 0 \\ 0 & λ < 0 \end{matrix}$ (3)

The recorded number T_v is used to remove noise from the image and it is also known as threshold point. T_v evaluation allows this method to escape the tedious selection of optimal parameter. The LBP operation does not describe degree of differences between the pixels. However this limitation is overcome in improved LBP method and is formulated as $ILBP = {\begin{matrix} 0 & T_{v} = 0 \\ \sum A_{v} (q_{j} - q_{c}) & 0 ⩽ T_{v} ⩽ 7 \\ 0 & T_{v} = 8 \end{matrix}$ (4)

where $A_{v} (q_{j} - q_{c)} = {\begin{matrix} A_{v} (q_{j} - q_{c}) & {if q}_{j} \geq q_{c} \\ 0 & otherwise \end{matrix}$ (5)

The purpose of recorded number T_v separates noise and put it in an easy area because T_v = 0 indicates an isolated noise while T_v = 8 represents an inaccessible point that produces noise in image. The gray scale value of every pixel is calculated using its surrounding pixels situated in input matrix. The middle pixel is subtracted from its surrounding pixels (Fig. 1). The surrounding pixels, whose values are less than or equal to middle pixel, are considered as ‘0’ else it is the difference between neighborhood and centre pixel. ILBP is obtained by adding these eight values.

This modification in LBP provides unequal results when the degree of difference between q_j and q_c is not same. Thus ILBP gives different values for matrices considered in Fig. 2 whereas LBP gives same values. Figure 1 shows the ILBP value of considered matrices.

4 Experimental results and discussion

The edge detection techniques discussed in the previous section are designed and simulated in MATLAB on Intel® Coretrademark i7 CPU @ 3.40 GHz, 4GB RAM PC. In this work two sets of comprehensive experiments are conducted to analyze the performance of proposed algorithm. Experiment-1 is conducted on synthetic images whereas experiment-2 detects edges for real images. Experiments are also carried out for compressed as well as noisy images. The threshold function T_v is used to remove residual noise and artifacts from image before edge detection. The performance of proposed edge detection algorithm is evaluated quantitatively with the help of accuracy. However the real life images taken from database do not have reference edge map therefore they are analyzed only in qualitative terms.

4.1 Experiment-1

The ILBP method is tested on synthetic image of size 138×110 (Fig. 3 column b) with and without blur. Figure 3 column (a) shows the standard deviation and column (c) shows the reference or actual edges of input image. The image is blurred by introducing Gaussian noise of different standard deviation values shown in column (d). The blurring effect is increased by increasing standard deviation. The columns (e)-(i) show the edge detection using Canny, Sobel, LBP, HLBP and ILBP methods respectively.

Fig.3

Effect of blur on edge detection of synthetic image. (a) Noise density, (b) Actual image, (c) Reference edges, (d) Blurred image, (e) Canny, (f) Sobel, (g) LBP, (h) HLBP and (i) ILBP segmented images.

It is observed from Fig. 3 that ILBP and Canny methods extract a continuous and complete edge map of all blurred images as compared to the other methods. The quantitative analysis of results obtained is given in Fig. 4 and Table 1. It is revealed that accuracy of proposed method is large and reduction in accuracy with increased blurring is also less as compared to other methods. It is also clear from the results that true positive (TP), true negative (TN) and accuracy of proposed method is significantly larger than other edge detection methods under consideration. Further false positive (FP) and false negative (FN) values are lesser than other edge detection methods except HLBP in false negative. Thus it is revealed that effect of blurring is quite less in case of ILBP as compared to other methods.

Fig.4

Accuracy, FP, FN, TN and TP for blurring effects in synthetic image.

Table 1

Performance measures of various segmentation methods for Gaussian noise in synthetic image.

Parameter\ND	1	2	3	4	5	7	9
Accuracy	ILBP	96.39	95.53	97.25	94.42	95.95	95.43	92.25
	HLBP	83.14	85.65	81.59	76.12	70.74	60.32	51.28
	LBP	73.36	70.37	70.68	69.64	68.19	64.90	62.07
	Canny	94.57	90.28	93.04	90.47	93.02	91.09	90.59
	Sobel	96.05	94.75	96.96	92.98	93.62	86.01	86.09
False Negative	ILBP	0.4606	0.5360	0.3874	0.6250	0.4876	0.4696	0.6239
	HLBP	0.1768	0.0011	0	0	0	0.0011	0.0034
	LBP	0.9977	1	1	1	1	1	1
	Canny	0.5191	0.7590	0.5282	0.7500	0.5405	0.5732	0.5800
	Sobel	0.4764	0.5957	0.4110	0.6137	0.5878	0.8851	0.9673
False Positive	ILBP	0.1565	0.2286	0.0822	0.3288	0.2050	0.3119	0.7005
	HLBP	2.7050	2.4527	3.1475	4.0822	5.0011	6.7827	8.3243
	LBP	5.1892	6.1655	6.0676	6.4403	6.9685	8.2399	9.4471
	Canny	0.4099	0.9032	0.6610	0.8795	0.6520	0.9493	1.0282
	Sobel	0.1982	0.3018	0.1092	0.5867	0.5023	1.5068	1.4110
True Negative	ILBP	15.9381	15.8660	16.0124	15.7658	15.8896	15.7827	15.3941
	HLBP	13.3896	13.6419	12.9471	12.0124	11.0935	9.3119	7.7703
	LBP	16.0439	16.0203	16.0405	16.0687	16.0788	16.0878	16.0935
	Canny	15.6847	15.1914	15.4336	15.2151	15.4426	15.1453	15.0664
	Sobel	15.8964	15.7928	15.9854	15.5079	15.5923	14.5878	14.6836
True Positive	ILBP	0.5394	0.4640	0.6126	0.3750	0.5124	0.5304	0.3761
	HLBP	0.8232	0.9989	1	1	1	0.9989	0.9966
	LBP	0.9955	0.9989	1	1	1	1	1
	Canny	0.4809	0.2410	0.4718	0.2500	0.4595	0.4268	0.4200
	Sobel	0.5236	0.4043	0.5890	0.3863	0.4122	0.1149	0.0327

^*Noise density (ND).

The effect of salt and pepper noise on edge detection is also analyzed on synthetic image. The algorithm is tested on noisy as well as clean images. The image is corrupted with salt and pepper noise of variable density. Figure 5 columns (a) to (d) shows the noise density value, the actual image, the noisy image and reference edge map respectively. Figure 5 columns (e) to (i) shows the detected edges using Canny, Sobel, LBP, HLBP, and ILBP methods respectively.

Fig.5

Effect of salt & pepper noise on edge detection of synthetic image. (a) Noise density, (b) Actual image, (c) Reference edge image, (d) Noisy image, (e) Canny, (f) Sobel, (g) LBP, (h) HLBP and (i) ILBP segmented images.

The comparison of edge detection techniques against reference image is given in Table 2 and Fig. 6. It is obvious from the results that accuracy of proposed method is close to 100% and variation in accuracy for different levels of noise is negligible.

Fig.6

Performance measure variations for salt & pepper noise in synthetic image.

Table 2

Performance measures of various segmentation methods for salt and pepper noise in synthetic image

Parameter\ND	0	0.05	0.10	0.15	0.25	0.30	0.40
Accuracy	ILBP	96.39	96.15	96.10	95.94	95.27	94.53	92.79
	HLBP	87.85	80.34	87.39	87.50	87.01	86.11	83.93
	LBP	73.36	72.84	72.35	71.65	70.90	70.12	69.50
	Canny	94.57	91.50	86.55	82.91	77.56	77.50	74.95
	Sobel	96.57	95.15	94.20	94.87	93.79	93.38	92.71
False Negative	ILBP	0.3360	0.4854	0.4831	0.4932	0.5304	0.5428	0.5777
	HLBP	0.1329	0.1982	0.1520	0.1622	0.1554	0.1836	0.2354
	LBP	0.9977	0.9977	0.9966	0.9989	0.9966	0.9977	0.9955
	Canny	0.7149	0.5158	0.5214	0.5270	0.5619	0.5394	0.5597
	Sobel	0.7126	0.4820	0.5450	0.6081	0.7444	0.8018	0.9144
False Positive	ILBP	0.1565	0.1723	0.1836	0.2016	0.3119	0.3930	0.6554
	HLBP	1.9448	3.1633	2.0034	1.9741	2.0653	2.1914	2.5113
	LBP	5.1892	5.3502	5.4955	5.7050	5.9550	6.1982	6.4032
	Canny	0.4099	0.9369	1.7782	2.3941	3.2748	3.3063	3.7218
	Sobel	0.1982	0.3468	0.4459	0.2680	0.3164	0.3300	0.3311
True Negative	ILBP	15.9381	15.9223	15.9110	15.8930	15.7827	15.7016	15.4392
	HLBP	14.1498	12.9313	14.0912	14.1205	14.0293	13.9032	13.5833
	LBP	16.0439	16.0428	16.0473	16.0315	16.0473	16.0541	16.0664
	Canny	15.6847	15.1577	14.3164	13.7005	12.8198	12.7883	12.3727
	Sobel	15.8964	15.7477	15.6486	15.8266	15.7782	15.7646	15.7635
True Positive	ILBP	0.5394	0.5146	0.5169	0.5068	0.4696	0.4572	0.4223
	HLBP	0.8671	0.8018	0.8480	0.8378	0.8446	0.8164	0.7646
	LBP	0.9955	0.9854	0.9381	0.9088	0.8930	0.8333	0.7894
	Canny	0.4809	0.4842	0.4786	0.4730	0.4381	0.4606	0.4403
	Sobel	0.5236	0.5180	0.4550	0.3919	0.2556	0.1982	0.0856

^*Noise density (ND).

The visual inspection of edges obtained using Canny method reveals that edges are more promising as compared to the proposed ILBP technique, however the close examination of results indicate superiority of ILBP on the basis of edge quality and accuracy. The close visual examination of results reveals that ILBP edges are more fine and continuous as compared to Canny edge map. In case of blurred synthetic image Canny performs slightly better for increased blurring, whereas for no blurring both the algorithms are equally good. In case of salt and pepper noise the performance of Canny method gets degraded for increase in noise level whereas ILBP finds edges more accurately.

The quantitative comparison using accuracy measure justifies the observations. The accuracy of Canny degrades by small amount for increased blurring whereas ILBP degradation is slightly more. However for good quality image ILBP accuracy is more. In case of salt and pepper noise the accuracy of ILBP is more and degradation of accuracy is less for increased noise level as compared to Canny method.

Thus ILBP method is better as compared to Canny, Sobel, LBP and HLBP methods for noisy images. It is indicated that proposed method overcomes the limitations of existing methods and detects the edges of digital images quite effectively in noisy environment.

4.2 Experiment - 2

Experiment 2 is designed to evaluate the performance of proposed ILBP edge detection technique on digital radiography images and real image data set. The real life and radiography images under consideration are taken from database. The database does not contain reference edge map therefore quantitative analysis is not carried out.

(a) Radiography Images

Different digital radiography pictures are considered to analyze the edge detection capabilities of various techniques. Digital radiography images are usually corrupted by noise and artifacts and thus require proper filtering so as to extract useful information from them. Segmentation helps in recovery of necessary information from these images. In this work, various images are considered for experimentation. However results of railway’s side frame and lower limb image are shown in the paper.

(1) Railway’s side frame

The radiography image under consideration is the side frame image of a railway, which shows the batch number and other parts. The experimentation is carried out on different sizes of image obtained by compression. Figure 7 represents the comparison of edge detection by different methods. Figure 7 column (a) and (b) show the image size and actual image respectively. Columns (c) to (g) show edge map using Canny, Sobel, LBP, HLBP and ILBP methods respectively. In large size image, Canny method extracted most of the edges except batch numbers. However in small images, it is not able to extract low contrast region and batch number. Further most of the edges and batch number are missed in all types of images using Sobel method. LBP method identifies fine changes, most of which are superfluous. HLBP extracted all edges however edges are thick and not clear. The ILBP method is able to extract the edges effectively for all compressed images and shows superior performance as compared to other methods.

Fig.7

Effect of size of image on edge detection of railway’s side frame. (a) Image Size, (b) Actual image, (c) Canny, (d) Sobel, (e) LBP, (f) HLBP and (g) ILBP segmented images.

(2) Lower limb Image

Digital radiography (DR) image of lower limb with broken tibia bone is considered for experimentation. Edge detection results of the DR image using Canny, Sobel, LBP, HLBP and ILBP methods are shown in Fig. 8. It is observed that all bones and skin edges are shown clearly and fracture of tibia bone is also visible using ILBP. The results obtained from other methods are not satisfactory because bones (like tibia, fibula, tarsal’s and metatarsals) and skin edges are not visible in the edge map. Thus ILBP provides a good edge map of lower limb digital radiography image.

Fig.8

Lower limb digital radiography image. (a) Original image, (b) Canny, (c) Sobel, (d) LBP, (e) HLBP and (f) ILBP segmented images.

(b) Real Image

The proposed method is tested on several real images out of which results of three real images of buffalo and two dogs are presented (Fig. 9). The implementation of proposed method on images with different scale is carried out for images of different class. This experiment is performed to show that a wide variety of real images with different resolution are analyzed in various applications. However performance analysis of the proposed edge detection technique is also carried out by considering an image with variable scaling so as to analyse the effect of image resolution on edge detection [38]. The size of the images are (237×131), (176×179) and (150×200) pixels. Figure 9 column (a) shows real images whereas the corresponding edge maps of different edge detection methods are given in Fig. 9 columns (b) to (f). In case of Canny method almost all edges are presented whereas some complex edges are missed. The Sobel edge map missed complex background edges. LBP method detected fine changes, most of which are unnecessary. HLBP extracted all the edges but they are thick and unclear. ILBP method extracted all the edges and is less affected by illumination and complex background. Thus ILBP segmented all three real images more effectively than the conventional methods.

Fig.9

Edge detection of real images. (a) Real images, (b) Canny, (c) Sobel, (d) LBP, (e) HLBP and (f) ILBP detected edges.

c) Effect of spatial resolution on edge detection

A parrot image is compressed using different scales to obtain (168×150), (134×120), (100×90) and (50×45) images (Fig. 10) with a view to compare and analyze the performance of developed algorithm under compression. The edges obtained using Canny method shown in Fig. 10 (b-e), are discontinuous and some edges are missed for highly compressed image. The Sobel edge map provides split edges and complex background is not detected. LBP method generates unnecessary edges whereas HLBP edges are thick and unclear. The edges obtained using ILBP are continuous, clear and accurately localized (Fig. 10 f). Thus ILBP edges are better as compared to other techniques under different spatial resolution conditions.

Fig.10

Edge detection techniques with different spatial resolution. (a) Real images, (b) Canny, (c) Sobel, (d) LBP, (e) HLBP and (f) ILBP detected edges.

In case of real and radiography images the performance of ILBP is much better as compared to Canny edge detection methods. It is observed that edges of ILBP are thin, continuous and accurate as compared to Canny and other methods. Further less prominent edges are also detected efficiently using ILBP whereas such edges are missed by Canny method. The above analysis proves that ILBP method besides rejecting blur and salt & pepper noise also detects the true edge map in digital radiography and real images.

The average computation time of designed methods for railway side frame of size (98×146) is given in Table 3. It is observed that LBP, Sobel and Canny methods are fast. LBP method has less mathematical calculation and thus low computational complexity whereas Sobel and Canny methods use first order derivatives. HLBP takes largest computational time. ILBP gives good edge detection but it takes large calculation time as compared to conventional edge detection methods. It is thus revealed that large computational time of suggested approach is acceptable because of its improved and robust performance in the facets of noisy and blurred conditions.

Table 3

Computational time for various edge detection techniques

Method	ILBP	HLBP	LBP	Canny	Sobel
Time (Sec.)	3.3657E-2	6.2397E-2	5.115E-3	25309E-2	3.435E-3
SD	3.12E-3	3.25E-3	2.1E-4	1.68E-3	1.2E-4

^*Standard Deviation (SD).

5 Conclusion

In this paper, a novel edge detection technique ILBP is proposed for noisy and compressed image by an improved counting scheme and local binary pattern. The performance of proposed method is evaluated on synthetic, medical as well as industrial digital radiography images and real-life images. The technique is tested on compressed, blurred and noisy images. The results show that ILBP is better than other edge detection techniques like Canny, Sobel, LBP and HLBP in terms of contrast invariance, edge discrimination capability and noise robustness. The edges detected using ILBP are continuous and strong in the presence of compression, blur and salt & pepper noise. Hence it is concluded from the analysis that ILBP outperforms the existing methods for edge detection of low-quality images.

References

Vardhana

, Arunkumar

, Lasrado

, Abdulhay

, Ramirez-Gonzalez

, Convolutional neural network for bio-medical image segmentation with hardware acceleration, Cognitive Systems Research 50 (2018), 10–14.

Yuan

, Zou

, Zhao

, Wang

, Hu

, Komodakis

, A robust and efficient approach to license plate detection, IEEE Transactions on Image Processing 26(3) (2017), 1102–1114.

Sarkar

, Venugopalan

, Reddy

, Ryde

, Jaitly

, Giering

, Deep learning for automated occlusion edge detection in RGB-D frames, Journal of Signal Processing Systems 88(2) (2017), 205–217.

Thru ADP014091, A., Defense Technical Information Centre Compilation Part Notice ADP014091.

Zhao

, Fu

, Sun

, Li

, Zhang

, Wang

, An effective method based on ACF for aircraft detection in remote sensing images, IEEE Geoscience and Remote Sensing Letters 14(5) (2017), 744–748.

Goulart

J.T.

, Bassani

R.A.

, Bassani

J.W.M.

, Application based on the Canny edge detection algorithm for recording contractions of isolated cardiac myocytes, Computers in Biology and Medicine 81 (2017), 106–110.

Sobel

, Camera models and machine perception, Stanford Univ Calif Dept of Computer Science (1970).

Prewitt

J.M.S.

, Object enhancement and extraction, Picture Analysis and Psychopictorics 10(1) (1970), 15–19.

Kirsch

, Computer determination of the constituent structure of biological images, Computers and Biomedical Research 4 (1971), 315–328.

10.

Verma

O.P.

, Hanmandlu

, Kumar

, Chhabra

, Jindal

, A novel bacterial foraging technique for edge detection, Pattern recognition letters 32(8) (2011), 1187–1196.

11.

Mendoza

, Melin

, Licea

, Interval type-2 fuzzy logic for edge detection in digital images, International Journal of Intelligent Systems 24(11) (2009), 1115–1133.

12.

Mendoza

, Melin

, Licea

, A hybrid approach for image recognition combining type-2 fuzzy logic, modular neural networks and the Sugeno integral, Information Sciences 179(13) (2009), 2078–2101.

13.

Melin

, Mendoza

, Castillo

, An improved method for edge detection based on interval type-2 fuzzy logic, Expert Systems with Applications 37(12) (2010), 8527–8535.

14.

Wang

P.Y.

, Wang

G.Y.

, Wang

, A fiber bundle geometry approach for edge detection of chromaticity distributions, IEEE Signal Processing Letters 24(11) (2017), 1691–1695.

15.

, Yuan

, Zhang

, Xiang

, Wang

, Zhang

, A dynamic mode decomposition based edge detection method for art images, IEEE Photonics Journal 9(6) (2017), 1–13.

16.

Deng

, Ma

, Yin

, An edge detection approach of image fusion based on improved sobel operator, International Congress on Image and Signal Processing (IEEE) 4 (2011), 1189–1193.

17.

Murala

, Jonathan Wu

Q.M.

, MRI and CT image indexing and retrieval using local mesh peak-valley edge patterns, Signal Processing: Image Communication 29 (2014), 400–409.

18.

Subrahmanyam

, Maheshwari

R.P.

, Balasubramanian

, Local maximum edge binary patterns: A new descriptor for image retrieval and object tracking, Signal Processing 92 (2012), 1467–1479.

19.

Yingchao

, Zhang

, Yin

, Ma

, Automatic detection of particle size distribution by image analysis based on local adaptive canny edge detection and modified circular Hough transform, Micron 106 (2018), 34–41.

20.

Lin

W.-C.

, Wang

J.-W.

, Edge detection in medical images with quasi high-pass filter based on local statistics, Biomedical Signal Processing and Control 39 (2018), 294–302.

21.

Biswas

, Hazra

, Robust edge detection based on Modified Moor Neighbour, Optik 168 (2018), 931–943.

22.

Guada

, Zarrazola

, Yáñez

, Rodríguez

J.T.

, Gómez

, Montero

, A novel edge detection algorithm based on a hierarchical graph-partition approach, Journal of Intelligent & Fuzzy Systems 34(3) (2018), 1875–1892.

23.

Chaira

, Fuzzy mathematical morphology using triangular operators and its application to images, Journal of Intelligent & Fuzzy Systems 28(5) (2015), 2269–2277.

24.

, Zeng

, Shen

, Jiang

, An effective edge extraction method using an improved local binary pattern for blurry digital radiography images, NDT & E International 53(1) (2013), 26–30.

25.

Gaidhane

V.H.

, Navdeep

A. Rani

, Singh

, An improved edge detection approach and its application in defect detection, In IOP Conference Series: Materials Science and Engineering 244(1) (2017), 012017.

26.

Rajathilagam

, Rangarajan

, Edge detection using G-lets based on matrix factorization by group representations, Pattern Recognition 67 (2017), 1–15.

27.

Moradi

, Mozaffari

, Hybrid approach for farsi/arabic text detection and localisation in video frames, IET Image Processing 7(2) (2013), 154–164.

28.

Tang

, Yin

, Sun

, Hu

, 3D face recognition using local binary patterns, Signal Processing 93(8) (2013), 2190–2198.

29.

Davarzani

, Yaghmaie

, Mozaffari

, Tapak

, Copy-move forgery detection using multiresolution local binary patterns, Forensic science international 231(1-3) (2013), 61–72.

30.

Zhu

, Bichot

C.E.

, Chen

, Image region description using orthogonal combination of local binary patterns enhanced with color information, Pattern Recognition 46(7) (2013), 1949–1963.

31.

Heikkilä

, Pietikäinen

, Schmid

, Description of interest regions with local binary patterns, Pattern Recognition 42(3) (2009), 425–436.

32.

Tan

X.Y.

, Triggs

, Enhanced local texture feature sets for face recognition under difficult lighting conditions, IEEE Transactions Image Process 19(6) (2010), 1635–1650.

33.

Nanni

, Lumini

, Brahnam

, Local binary patterns variants as texture descriptors for medical image analysis, Artificial Intelligence in Medicine 49 (2010), 117–125.

34.

Qian

, Hua

X.S.

, Chen

, Ke

, PLBP: An effective local binary patterns texture descriptor with pyramid representation, Pattern Recognition 44(10-11) (2011), 2502–2515.

35.

Liao

, Law

M.W.K.

, Chung

A.C.S.

, Dominant local binary patterns for texture classification, IEEE Transactions on Image Processing 18(5) (2009), 1107–1118.

36.

Guo

, Zhang

, Rotation invariant texture classification using LBP variance (LBPV) with global matching, Pattern Recognition 43(3) (2010), 706–719.

37.

Zhu

C.R.

, Wang

R.S.

, Local multiple patterns based multi-resolution gray-scale and rotation invariant texture classification, Information Sciences 187(1) (2012), 93–108.

38.

Parah

S.A.

, Sheikh

J.A.

, Akhoon

J.A.

, Loan

N.A.

, Bhat

G.M.

, Information hiding in edges: A high capacity information hiding technique using hybrid edge detection, Multimedia Tools and Applications 77(1) (2018), 185–207.