A Study of Low-resolution Safety Helmet Image Recognition Combining Statistical Features with Artificial Neural Network

Transcription

1 A Study o Low-resolution Saety Helmet Image Recognition Combining Statistical Features with Artiicial Neural Network Xinhua JIANG, Heru XUE *, Lina ZHANG, Yanqing ZHOU College o Computer and Inormation Engineering, Inner Mongolia Agricultural University, Hohhot, Inner Mongolia 0008, China College o Physics and Electronic Inormation, Inner Mongolia Normal University, Hohhot, Inner Mongolia 000, China Abstract According to the entrance and exit to building sites monitoring video, this paper discusses the recognition method o the saety helmet o the low-resolution image captured rom video and deduces the relationship between dierent eatures and recognition rate taking account o the low-resolution saety helmet recognition problem at the long distance. It irst captured heads in the monitoring video and then extracted the statistical eatures with Local Binary Pattern and gray level co-occurrence matrix. Finally, the recognition rate o test sample was calculated by taking advantage o Back-Propagation artiicial neural network. The experimental results show that the GLCM statistical eatures in combination with BP artiicial neural network can recognize the saety helmet eectively. The recognition rate reaches 94%. So the method proposed by this paper is ast and easible. (Abstract) Keywords - Low-resolution; helmets; LBP; GLCM; artiicial neural network (key words) I. INTRODUCTION The saety helmet, a vital protective measure to ensure the saety o the sta, is the most commonly used and the most eective protect means in the workplace. It ever has saved countless lives. The statistics o occupational injuries rom United States in 995 showed that the number o deaths caused by not wearing the saety helmet accounted or a quarter o all the deaths, so we can understand the importance o wearing the saety helmet []. By the image processing technique, how to identiy workers in images who wear the saety helmet has important realistic meaning and has become a signiicant research direction in the ield o image recognition []. For example, the Central Police University in 004 put orward a circle/circular arc detection method based upon the modiied Hough transorm, and applied it to the detection o saety helmets or the surveillance system o ATMs [3]. The China University o Mining and Technology in 0 constructed standard images o helmet, extracted our directional eatures, modeled the distribution o these eatures using a Gaussian unction and separated local images o rames into helmet and non-helmet classes. Finally, the detection rate was 83.7% [4][5]. In other universities, students researched the helmet presence classiication with motorcycle detection and tracking [6]. The video o the entrance o building sites are oten aected by the complex background, uneven lighting and camera position. In general, because o the limitation o data storage space, network transmission speed and image acquisition equipment actors, the head and ace images are low-resolution [7]. There were some researches o lowresolution based on ace and ear recognition and so on [7][8]. These techniques have relatively matured. However, in view o the low-resolution saety helmet image recognition, the above mentioned actors make it more diicult to detect and recognize the saety helmet in building site videos and cause the results o traditional methods to be unsatisactory. This paper ocuses on both the eature extraction and recognition or the low-resolution saety helmet. The ways o image eature extraction adopt Local Binary Pattern (LBP), Gray level co-occurrence matrix (GLCM), and Back- Propagation (BP) artiicial neural network are used to classiy. Finally, we propose a recognition method based upon the Haralick texture statistical eatures, and apply it to the BP artiicial neural network to train data and recognize the testing data. II. MATERIAL AND PREPROCESSING A. Material Images which the experiment used were captured rom the monitoring videos o entrance and exit to building sites, and the monitoring videos were compressed by video server and transmitted to a personal computer through long distance. The monitoring cameras are usually installed ar rom the entrance and at a higher place in order to avoid covered rom other objects. So by remote transmission, the videos we have got are low-resolution. The captured video is 4 rames per second and the size is pixels. In the experiment, because workers who located in the edge o video images can not be seen clearly, the captured region is the center area o video images. The captured window scope is pixels. The head region o workers was captured by using the OpenCV every 5 rame, and normalized to pixels. 78 head images were collected rom building site monitoring videos. In general, a saety helmet is maybe red, white, blue or yellow. Sometimes some women heads are covered by the hood. As a consequence, the pure color is DOI 0.503/IJSSST.a ISSN: x online, print

2 diicult to distinguish whether workers wear saety helmets or not. Preprocessing o saety helmet image beore eature extraction is essential. At the same time, because o the light and noise inluence, original images are very hard to extract eatures. Eventually, the collected images o the video in igure are shown as igure. Figure. Original Video Frames The igure 3 shows an example o the LBP operator. The operator assigns a label to every pixel o an image by threshold the 3 3 neighborhood o each pixel with the center pixel and considering the result as a binary number. In act, let gc and g0,,g7 denote respectively the gray values o the center and its eight-neighbor pixels, then the LBP code or the center pixel with coordinate (x) is calculated by Eq. (). 7 p LBP ( x ) s ( g c g ) () p 0 Where s(z) is the threshold unction. When z is not less than 0, the value o s(z) is. When z is less than 0, the value o s(z) is 0 []. The law o the basic LBP operator is that it covers only a small area within a ixed radius. Apparently, it is not suicient or dierent sizes and requencies o the texture. Soon aterwards, the LBP operator is extended and improved to use neighborhoods o dierent sizes by Ojala T [9] and other people. The circle replaces the cube. In the special local neighborhood, we can obtain the any radius and sample points. As long as a sample point does not all into the center o a pixel, bilinear interpolation is adopted. In the ollowing, the notation (P, R) will be stand or pixel neighborhoods, which means P sampling points on a circle o radius o R, See igure 4 or an example o the circular(8, ) neighborhood [][3]. p Figure. The Head Area o the Captured Images B. LBP Operator Image preprocessing is convenient to the next section, which carries on converting rom the color to gray image. When using the grid extracts eatures, the size o images is set as 0 0 pixels. The LBP operator, introduced by Ojala T [9], serves or texture representation. Now, the LBP is widely applied in various computer vision ields and is regarded as a complementary measure or image local contrast. However, the eature extraction is one o the key technologies o image recognition. In the recent years, LBP algorithm has been successully applied to the eature extraction o low-resolution images. The LBP has the robust classiication ability, high computing eiciency and less susceptible to illumination variance. In addition, LBP operator has invariance or monotonous gray level change. So it has achieved excellent result in the saety helmet recognition o uneven illumination [0]. Figure 4. The Circular (8,) Neighborhood When using the circular neighborhood, the patterns o P types generated by LBP (P, R) are seen as a negative or texture extraction, recognition, classiication and access to inormation [4]. When the parameters o LBP(R, P) are P=8 and R=, hence, there are P=8=56 sub-binary numbers to describe texture eature. So the uniorm patterns are put orward to reduce its dimensionality. A Local Binary Pattern is called uniorm i it contains at most two bitwise transitions rom 0 to or vice versa when the binary string is considered circular. For example, the patterns (0 transitions), (a) The Original Images (b) The Processed Images Ater LBP Figure 3. The Calculation Method o the Basic LBP Operator Figure 5. The Processed Images With LBP DOI 0.503/IJSSST.a ISSN: x online, print

3 00000( transitions) and 00( transitions) are uniorm whereas the patterns 0000(4 transitions) and 0000(6 transitions) are not. In the computation o the LBP histogram, uniorm patterns are used so that the histogram has a separate bin or every uniorm pattern and all non-uniorm patterns are assigned to a single bin[3]. Due to the complexity o images, uniorm patterns accounts or a bit less on the all kinds o images. Aterwards, Maenpaa [5] extended the LBP operator again and mentioned the LBP operator based on the rotation invariant. Images used the LBP operator processing is shown in the igure 5. III. EXTRACTING FEATURES A. Gray level co-occurrence matrix Texture eature analyses have been widely employed in classiication with proper eature selection, where classiier design can be greatly simpliied. Hence we adopt texture eatures rom GLCM [6] to identiy with helmet. The GLCM is a pixel-based image processing method. An element P (i, j, d, φ) o the GLCM o an image represents the relative requency, where i is the gray level at location (x), and j represents the gray level o a neighboring pixel at a distance d and an orientation φ rom location (x). The creation o the GLCM matrix is based on the distance between pixels, the pixels angle(0,45,90,35 ) and the number o level grayscale conversion (maximum is 56) parameters [7]. Presume (x) as a two-dimensional digital image, its size is M N, gray level is Ng, and then the GLCM depending on the angle and distance parameters, is expressed in Eq.() [ 6,8, 9] : P j, d, ) # {( x ), ( x ) M N () ( x ) i, ( x ) j} Here, #(x) expresses the number o elements in the collection and P is an M N matrix. I the distance between x,y and x,y is d, the angle between the latter two and the axis is φ ou can get the GLCM o the distance and angle P(i, j, d, φ ) [9]. The content inormation between two neighboring cells {(x), (x)} separated by a distance d and an orientation φ is represented by P (i, j, d, φ ). The texture eatures (contrast, correlation, entropy, energy, and homogeneity) used in this study. ) Contrast This statistic measures the spatial requency o an image and is dierence moment o GLCM. It is the dierence between the highest and the lowest values o a contiguous set o pixels, so higher contrast values indicate large local variations: N g m { Pd, }, i j m 0 i j m (3) The corresponding value or the coarse texture is smaller, and the corresponding value o a ine texture is larger [9]. ) Correlation The correlation eature is a measure o gray tone linear dependencies in the image. Being derived rom the GLCM, correlation is a two-dimensional requency histogram in which individual pixel pairs are assigned to each other on the basis o a speciic, predeined displacement vector [0] : ( i Pd, j) x y i j (4) x y 3) Entropy Entropy measures disorder o the image and also indicates complexity within an image. When the value o Pd, φ (i, j) is less and more dispersed, entropy is larger; on the other hand in case o a higher concentration, the entropy value is small. It is highly correlated to energy. Non-uniorm texture has a high entropy value [] : 3 P, log Pd, i j d 4) Energy Energy expression the repetition o pixel pairs o an image and also it is a measure o homogeneity o an image. It is the measure o image gray-level distribution, when the value distribution o Pd, φ (i, j) is more concentrated on the main diagonal, the value o energy is relatively large; otherwise the value is small: 4 P d, i j (5) (6) 5) Homogeneity Homogeneity is inversely proportional to contrast at constant energy whereas it is inversely proportional to energy []. It deines the uniormity o the distribution o elements in the GLCM to its diagonal. That means, i the elements o GLCM are more diagonally distributed, then homogeneity is high, as i-j is less [0]. 5 i j P d, ( i B. Artiicial neural network The artiicial neural network (ANN) method is based on the indings o a biological nervous system []. In this artiicial neural system, there are nerve cells which are joined together in a variety o ways to orm networks. The ANN consists o layers, namely the input layer, the hidden layers and the output layer. The hidden layers in-between these layers is where the data are processed. The number o the nerve cells in hidden layer is signiicant or the perormance as well as the length o the network. Usually, the output o each neuron is determined by using a nonlinear activation unction such as a sigmoid and hyperbolic tangent. ANNs are trained by experience, when applied to an unknown input in the network; it can generalize rom past experiences and produce n new result. The igure 6 shows a undamental representation o an artiicial neuron network. (7) DOI 0.503/IJSSST.a ISSN: x online, print

4 Figure 6. The Fundamental Representation o Artiicial Neuron Network The inputs o the network are represented by symbols x, x,..., x n. Each o these inputs is multiplied by a connection weight that is represented by o the w, w,..., w n. In ANN, x w, x w,..., x n w n results are summed and ed through an activation unction to generate a result or output. θi is a bias value, G(x) is an activation unction. IV. PROPOSED METHOD FOR LOW-RESOLUTION SAFETY HELMET IDENTIFICATION In this study, a method based on the GLCM and ANN methods were used or identiication o saety helmets. A total o 78 images were used, belonging to classiications having saety helmet or not having saety helmet. Each o the images was cropped to a pixel image beore processing. In this study, ive texture eatures o images were extracted. Each texture eature is computed or dierent GLCM(orientations are 0, 45, 90, 35 ) and distances (d=,, 3, 4) parameters that were established rom the LBP image o the images. The average value o the texture eatures were computed and used as the input nodes in the ANN. The study was consisted 4 parts, and shown in igure 7. Figure 7. The Proposed Method For Classiication The processes can be summarized as ollows: The Image Preprocessing: The obtaining o 78 images o saety helmets. In this part, a preprocessing and essential step is implemented, which includes resizing the images and LBP preprocessing. The GLCM Features: Obtaining data set consisting o ive texture eatures o 78 images. The texture eature o GLCM was computed or dierent orientations (0, 45, 90, 35 ) and then the average value o the texture eatures was computed and used as the input nodes in ANN. The dimensions here describe dierent eatures resulting rom the GLCM, the total size o data set is 5 556, where 5 is the dimension o eature size o each images and 556 comes rom 78 samples per class multiplied by classes. ANN:In this study, the texture eatures are applied as input to ANN, so there are 5 nodes in input layer. The number o nodes o hidden layer is 30 or 40. The output is having saety helmets or not in images. The training accuracy o network is 0.00, the maximum iteration times are 5000, the learning rate is 0.0. The weights and thresholds are adjusted by the gradient descent method. The output result o network are 0 and, where means wearing helmet and 0 means without wearing helmet. V. CONCLUSIONS The aim o the study was to obtain an accurate identiication o wearing helmets or not by texture eatures. The textures are seen in igure 5. The texture (contrast, correlation, entropy, energy and homogeneity) were used or classiication o wearing helmets or not through ANN. The distribution o textures is given in igure 8, 9 and 0. The Camera position, light intensity, helmets shooting site, the video image resolution, helmet and clothing color dierence, human skin change with a great range. It is to such an extent that these eatures play an important role in the distinction o wearing helmets or not. While these kinds DOI 0.503/IJSSST.a ISSN: x online, print

5 o eatures are seen as classiied characteristics as long as being limited in a scenario, sometimes in distinction o wearing helmets or not are very alike. In recent years, various techniques are used in low resolution video recognition, it is seen that machine learning techniques were not among them. In this study, LBP, which is an image reprocessing technique, GLCM, which is an image character Figure 0. Distribution o Energy, Entropy and Homogeneity Features Figure 8. Distribution o Correlation, Energy and Entropy Features description technique, and ANN, which is a machine learning technique, are used in identiication o wearing helmets or not. A total o 78 images belonging to categories were used in this study. Classiication processes with ANN was carried out or dierent training-test rates o images in Table. 94% classiication rate was achieved or the 70%-30% training-test partition. According to the results, it was seen that texture eatures are import parameters in classiication o wearing helmets or not o video images. TABLE I. THE PERFORMANCES OF GLCM AND ANN IN IDENTIFICATION WEARING HELMETS OR NOT The proportion o training-test Training set Test set Correctly classiied images Classiication accuracy 30%-70% % 50%-50% % 70%-30% % Figure 9. Distribution o Correlation, Entropy and Homogeneity Features To show the eectiveness o the GLCM texture eatures in LBP images, a comparison with histogram o LBP images was carried. In the paper, combining LBP with histogram statistics characteristics, and adopting the BP ANN as classiier, and setting the hidden layer nodes as 0, 0, and 40 respectively, and using elastic gradient descent method, the right classiication rate is shown in Table. TABLE II. PERFORMANCE VALUES FOR HISTOGRAM AND GLCM FEATURES Feature Extraction Method LBP+Histogram +BP LBP +GLCM +BP Test Images Correctly Classiied Images Accuracy % % According to the results, texture eatures o low resolution images proved their useulness in identiication o wearing DOI 0.503/IJSSST.a ISSN: x online, print

6 helmets or not. One o the major reasons o this success was that the eature extraction with combining LBP and GLCM algorithm were used in classiication o low resolution images. ACKNOWLEDGEMENTS This work was supported by natural science oundation o Inner Mongolia autonomous region (No.04MS063), and supported by national natural science oundation o China (No.64604, No.64604). REFERENCES [] Maybank S and Tan T. Introduction-special section on visual surveillance, International Journal o Computer Vision, vol.37,no., pp.73-73, 000 [] Hu Tian, Wang Xingang. Saety Helmet Recognition based on Wavelet Transorm and Neural Network System Analysis and Design, Sotware Guide, vol.3, pp.37-38, 006 [3] C-Y Wen. The Saety Helmet Detection Technology and Its Application to the Surveillance System, Journal o Forensic Sciences, vol.49, No.4, pp., 004. [4] Cai Limei, Qian Jiansheng. A method or detecting miners based on helmets detection in underground coal mine videos, Mining Science and Technology, vol.,no.4, pp , February 0. [5] Cai Limei. Research on human detection and tracking in underground coal mine videos, Ph.D. thesis, School o Inormation and Electrical Engineering, ON,Jiangsu,pp.56-64,00. [6] J. Chiverton. Helmet presence classiication with motorcycle detection and tracking. The Institution o Engineering and Technology, Vol.6, No.3, pp.59 69, 0. [7] Zhou Yi. Research o low-resolution ace recognition technology,m.s. thesis, University o Electronic Science and Technology o China, ON Chengdou, pp. -6, 0. [8] Wang Xiaoyun, Yuan Weiqi, Guo Jinyu, Study o low resolution human ear image recognition, Application Research o Computers, vol.7,no.,pp , November 00. [9] T. Ojala,M. Pietikainen,T. Maenpaa. Multiresoluti-on gray scale and rotation invariant texture classiication with local binary pattrerns. IEEE Trans. on Pattern Analysis and Machine Intelligence,vol.4,No.7, pp , 00. [0] Zhang Zheng, Wang Yanping, Xue Guixiang. Digital image processing and machine vision Visual C++ and Matlab, CA: Beijing, Posts & Telecom Press, pp ,00. [] Yang Bo, Chen Songcan. A comparative study on local binary pattern (LBP) based ace recognition: LBP histogram versus LBP image, Neurocomputing, vol.0,no.3, pp: ,november 03. [] Timo Ahonen, Abdenour Hadid, Matti Pietikäinen, Face Description with Local Binary Patterns: Application to Face Recognition, IEEE Transactions on Pattern Analysis and Machine Intelligence, vol.8,no., pp:037-04,january 007. [3] Timo Ahonen, Abdenour Hadid, and Matti Pietikäinen. Face recognition with local binary patterns, ECCV, pp , 004. [4] Xu Tao. Research on Face Recognition based on local binary pattern operator, M.S. thesis, School o Inormation Science & Engineering Central South University, ON Hunan, pp.7-0, 0. [5] T.Maenpaa, M.Pietikainen., Texture analysis with local binary patterns. Handbook o pattern recognition and computer vision. 3rd ed., World Scientiic, Singapore, pp.7-736, 00. [6] Haralick, R.M.,Shanmugam, K.,Dinstein, Its'Hak. Textural eatures or image classiication. IEEE Trans. Syst. Man Cybern. Vol.3,No. 6, 60 6,973 [7] SN Ondimu,H Murase. Eect o probability-distance based Markovian texture extraction on discrimination in biological imaging. Comput. Electron. Agric. Vol.63,,008 [8] Guang-ming Xian. An identiication method o malignant and benign liver tumors rom ultrasonography based on GLCM texture eatures and uzzy SVM. Expert Syst. Appl. Vol.37, ,00 [9] Shi Yujing,Bai Haijing, Wang Xuejun. Image classiication o education resources based on texture eatures. Procedia Engineering, vol.9, ,0 [0] S.Dutta,A.Datta,N.Das Chakladar, S.K.Pal, S.Mukhopadhyay, R.Sen. Detection o tool condition rom the turned surace images using an accurate grey level co-occurrence technique. Precision. Engineering. Vol.36, ,0 [] K.Manivannan, P.Aggarwal, V.Devabhaktuni, A.Kumar, D.Nimsb, P.Bhattacharya. Particulate matter characterization by gray level cooccurrence matrix based support vector machines. Journal o Hazardous Material. Vol.3 4, 94 03,0 [] Lynne Boddy, Colin W. Morris. Artiicial Neural Networks or Pattern Recognition. Machine Learning Methods or Ecological Applications, pp.37-87,999. DOI 0.503/IJSSST.a ISSN: x online, print