محاسبه الگوی اصلی دودویی محلی برای تشخیص چهره با استفاده از ابزارهای داده کاوی
|کد مقاله||سال انتشار||مقاله انگلیسی||ترجمه فارسی||تعداد کلمات|
|22272||2012||7 صفحه PDF||سفارش دهید||5600 کلمه|
Publisher : Elsevier - Science Direct (الزویر - ساینس دایرکت)
Journal : Expert Systems with Applications, Volume 39, Issue 8, 15 June 2012, Pages 7165–7172
Local Binary Patterns are considered as one of the texture descriptors with better results; they employ a statistical feature extraction by means of the binarization of the neighborhood of every image pixel with a local threshold determined by the central pixel. The idea of using Local Binary Patterns for face description is motivated by the fact that faces can be seen as a composition of micro-patterns which are properly described by this operator and, consequently, it has become a very popular technique in recent years. In this work, we show a method to calculate the most important or Principal Local Binary Patterns for recognizing faces. To do this, the attribute evaluator algorithm of the data mining tool Weka is used. Furthermore, since we assume that each face region has a different influence on the recognition process, we have designed a 9-region mask and obtained a set of optimized weights for this mask by means of the data mining tool RapidMiner. Our proposal was tested with the FERET database and obtained a recognition rate varying between 90% and 94% when using only 9 uniform Principal Local Binary Patterns, for a database of 843 individuals; thus, we have reduced both the dimension of the feature vectors needed for completing the recognition tasks and the processing time required to compare all the faces in the database.
Nowadays, the use of cryptographic techniques has become an essential requirement for a huge amount of applications. In most cases, the weakness of security systems still remains on the verification of the identity of the remote user. In relation to this, in the last 15 years there has been an intensive research to develop new security systems, much of them involving the so-called biometric features (Bowyer et al., 2008, Hao et al., 2006 and Zhao et al., 2003). A great number of theoretical and experimental innovations have made automated biometric authentication not only technically feasible, but also extremely popular. Automated biometric systems are being widely used in many applications such as surveillance, digital libraries, forensic work, law enforcement, human computer intelligent interaction, and banking, among others. For applications requiring high levels of security, biometrics can be integrated with other authentication tools, such as smart cards and passwords. In general, a biometric system is a pattern recognition system that makes a personal identification by determining the authenticity of a specific physiological or behavioral characteristic possessed by the user. One of the most popular biometric techniques to identify users is face recognition. Thus, an automated face recognition system is commonly based on extracting a set of features from the user’s face, either geometric characteristics or some information of textures and shapes of the different elements in a human face (Li & Jain, 2005). Among others, the main advantages of using facial recognition systems include their reliability, providing non-collaborative user recognition and incorporating some additional information, such as facial expressions. As a consequence, it is an emerging research area and, in the next few years, it is supposed to be extensively used for automatic human recognition systems in many different areas, such as access control, human–machine interfaces, robotics, etc. Face recognition systems are often classified depending on the method used to obtain the face features. On the one hand, holistic methods use the whole face region as the raw input; they include well-known techniques, such as Principal Components Analysis (PCA) (Turk & Pentland, 1991) or Linear Discriminant Analysis (LDA) (Etemad & Chellappa, 1997), among others. On the other hand, local or feature-based methods extract facial features, such as eyes, nose and mouth; their locations and local statistics are the input to the recognition stage. Some of most used algorithms for local feature analysis are Elastic Bunch Graph Matching (EBGM) (Wiskott, Fellous, Kuiger, & von der Malsburg, 1997) and, more recently, Local Binary Patterns (LBP) (Ahonen et al., 2004 and Ojala et al., 1996). Local Binary Patterns (LBP) were introduced as a robust descriptor of microstructures in images (Ojala et al., 1996). The main advantages of this operator are its tolerance to illumination changes, its computational simplicity and its invariance against changes in gray levels. Ahonen et al. (2004) published a novel approach: in their study they present the LBP operator as a powerful texture descriptor applicable to face recognition. This method involves dividing the face image into regions and applying the LBP operator in all of them. Once the operator has been applied, the histogram of each region is computed. Since each face region contributes in a different way for the recognition process, a specific weight is assigned to each histogram. Afterwards, the similarity distance between two faces is obtained by comparing the histograms of each region, and the recognition process is then completed. Although the original algorithm had a high recognition rate (around 94%), this method presents two main drawbacks: it needs a large feature vector and the assignment of the specific weights to each part of the face was not much precise. As a consequence, our work firstly aims at reducing the dimensions of the feature vectors required to perform the face recognition process using the LBP operator. To do this, we propose the use of what we call the Principal Local Binary Patterns, which include the most relevant information required to recognize a face; they will be calculated by means of data mining tools. On the other hand, we have also designed a 9-region mask, belonging to the most important face areas related to face recognition, such as the eyes, the nose or the mouth; then, a proposal for calculating the weights assigned to each face region – using data mining tools, as well – will be also introduced. The experiments show that our method gives accurate results and that using data mining tools allows to reduce the feature vector necessary to compare the faces in the database without affecting the final recognition rate. The paper is organized as follows: firstly, Section 2 presents the LBP operator and some recent extensions to the original algorithm. Then, the method designed for computing the Principal Local Binary Patterns and for assigning the weights to each face region is described in Section 3. After that, Section 4 shows the results of the experiments implemented on the FERET database and compares them to previous works. Finally, some concluding remarks and future works are shown in Section 5.
نتیجه گیری انگلیسی
Local Binary Patterns have proved to be a powerful measure of image texture, showing excellent results in terms of accuracy. The original LBP operator had two main problems: the computational complexity needed to describe a face with the operator, and that some weights were not chosen in a precise way. In this work, we have presented a method to obtain the Principal Local Binary Patterns required to characterize a face image and, then, to assign the region weights properly, using in both cases data mining tools. As a result, we have significantly decreased (by 83.05%) the dimension of the histograms containing the information of a face, achieving a recognition rate of 94%. Moreover, we have designed a mask with 9 regions and obtained a set of optimized weights for this mask; using the 9-region mask, a recognition rate of 90% is obtained, reducing the length of the original feature vector by 96.09%. Future works aim to design further methods to obtain higher recognition rates with small feature vectors. Consequently, it would be desirable to know more precisely the importance of each area of the face during the recognition process; thus, the specific weights of each region could be optimally set. On the other hand, most of the misclassified images in our experiments show two patterns: drastic changes in the expression or great variations in pose. Therefore, we are currently working to improve the results and finding a solution for this kind of problems