Dissertation for Master's Degree in Computer Science
Abstract
Nowadays, in many fields, we need devices that identify people's computers and recognize them based on their body characteristics. Face recognition system as a biometric system is basically a pattern recognition system that recognizes a person based on the vector of certain physiological characteristics or behavior. After extraction, the feature vector is usually stored in the database. The main purpose of this research is to study and investigate the effect of choosing the appropriate features of images using the cuckoo search algorithm. Therefore, choosing an optimal subset according to the large dimensions of the image feature vector to speed up the face recognition algorithm can be necessary and important. First, we extracted the features of face images from the existing database, then we selected an optimal subset of face features by using binary cuckoo algorithm. This subset of optimal features was evaluated by K-nearest neighbor classifiers and neural networks, and by calculating the classification accuracy, it was observed that the proposed method is able to recognize faces with an accuracy of over 90% based on the important features selected by the proposed algorithm. Keywords: face recognition, cuckoo algorithm, feature extraction, feature selection.
Introduction
We all know that a series of features are used to identify each other, which is exclusive to each person and differs from person to person. Today, in many fields, we need tools that identify the identity of people and recognize them based on their body characteristics, and this field is growing more and more every day and has found many interested parties.
Recent advances in information and imaging technologies allowed biometric systems to be developed. Biometric systems are able to extract an identity recognition sample and compare it to reference data to show whether the claimed identity is matched by the individual or not. Various identifiers are used for identity recognition, including fingers, hands, feet, faces (facial images), eyes, ears, teeth, veins, voices, signature (style), typing, walking. Face recognition has advantages over other biometric methods such as fingerprint recognition. In addition to the naturalness and inevitability of this type of recognition, the most important advantage of face recognition is that the face can be captured and covered at any distance. Various techniques have been developed for face recognition. These techniques have different goals according to the type of face recognition. In some applications, recognition speed is important, and in some cases, these techniques seek to increase the speed and accuracy of recognition. 1-2 Face recognition is one of the most important approaches in the field of machine vision People from a specific database has been the focus of any automated security system without the need for a user for years. The applications of these systems are expanding over time and with increasing efforts to improve security systems.
The first and most important step to identify any face is to find the face in an image, because if there is an error in this step, the next steps will not be executed correctly. In security systems (identity verification), the location of the face can be assumed to be fixed, that is, it is required that the face be placed in a predetermined location with the same lighting settings. In this way, there is no need to find a person's face in an image anymore, and by removing fixed parts from the input image, the face can be delivered to the face recognition system as input. Therefore, finding a face in an image is beyond a face recognition system, for example, in algorithms for detecting the number of people in an image, alarm systems, encoding video images for low transmission speed, are also used.
Another approach is face detection from the point of view of pattern recognition [1]. One of the problems of pattern recognition is finding a face in an image. The face as a pattern to be recognized is very complex.The face can have different expressions. A single face has distinct and recognizable expressions from another, such as expressions of happiness, sadness, and surprise, a part of the face may be covered with objects such as a hat or glasses, and it may also have different orientations and rotations. According to these cases, face detection in the field of pattern recognition is one of those problems for which a clear solution has not been provided so far, and many researches are still being conducted in this field. Face detection is very important, because in systems such as face tracking systems and face recognition systems, the first and most basic task is to find the location of the face. So far, many methods have been used to find faces. 1-3 issues During recent years, the demand for creating security through biometric methods [2] has become extremely popular. The use of various biological data such as fingerprints, handwriting, face and iris is presented. Among the various biometric methods, the use of facial images is more easy to capture than other methods due to its non-intrusive nature, and therefore has a wider range than other methods. For better face recognition systems, it is necessary to store images of people with different angles in the database, so with the increase of images in the database, their processing speed decreases and the storage space also decreases. On the other hand, face recognition is one of the most important applications from a computer point of view, especially for surveillance. This means that face recognition and its components is one of the most important approaches to facilitate the communication between man and machine and the active areas in the field of machine vision science.
The first step in the face recognition system is face recognition, for this reason, in order to recognize and extract the face region from non-faces, first the features of the face and non-face must be extracted and finally the face recognition should take place. In this way, the performance of identification methods based on two-dimensional face images is affected by environmental conditions, and things such as changing the intensity of ambient lighting, changing the angle of light radiation, changing the angle of the face, facial expression, changing age, etc. affect the performance of these methods. Therefore, the problem of face recognition has been investigated by researchers for many years and many classification algorithms have been proposed for it. In fact, the ultimate goal of a classified algorithm is to achieve a pattern recognition system and achieve the highest possible classification rate for the given problem. It is not far from the mind that none of these methods alone can solve this problem, and each classification has its own strengths and weaknesses.
In face recognition systems, with the increase in the database of face images, there is a problem of speed reduction and lack of memory. To solve these problems, the modified cuckoo search algorithm can be used in feature selection. Face recognition is usually used under the supervised learning model, in which the system is trained to perform recognition operations using sets of training examples. After the training period, a new face is submitted to the system to be recognized based on the existing patterns.
The purpose of feature selection is to reduce the size of the problem and reduce the search space for results for learning algorithms. The optimal selection of a subset of M elements among N features F=[Amax f1,f2,f3,…….fn], so that M < N and is optimal compared to other subsets of the same size, and at the same time increases the accuracy of prediction. In classifier design for pattern recognition, feature selection can increase the accuracy and speed of the classifier. Therefore, in recent years, various researches have been proposed in the field of feature selection, cuckoo search algorithms (COA), ant colony optimization algorithm (ACO), genetic algorithm (GA), particle group optimization algorithm (PSO) and colonial competition algorithm (ICA) have received more attention. These algorithms use many iterations and experiences from previous iterations to reach the optimal solution. Apart from the methods mentioned above for feature selection, which are mainly known as wrapper methods, there is another group of methods known as filter methods. In this method, no classification function is used to evaluate the selected subsets.