Age Estimation And Gender Classification Based On Face Detection And Feature Extraction

Nowadays the computer systems created a various types of automated applications in personal identification like biometrics, face recognition techniques. Face verification has turn into an area of dynamic research and the applications are important in law enforcement because it can be done without involving the subject. Still, the influence of age estimation on face verification become a challenge to decide the similarity of pair images from individual faces considering very limited of data base availability. We focus on the development of image processing and face detection on face verification system by improving the quality of image quality. The main objective of the system is to compare the image with the reference images stored as templates in the database and to determine the age and gender.


INTRODUCTION
The importance of this paper is the design of image processing system and face recognition on face verification system to improve image quality with the purpose of identify the level similarity of face images based on the age stages and finding the gender of the persons. The first step is Image processing where quality of face image is improved and enhanced using histogram equalization methods. Brightness Preserving Dynamic Fuzzy Histogram Equalization (BPDFHE) method [3] is used for increasing brightness and contrast enhancement and translation of color images into grayscale to get accurate result. Histogram equalization is mainly useful in images with foregrounds and backgrounds that are both dark or both bright. Especially, the method can direct to better outlooks of bone structure in x-ray images, and to better detail in photographs that are over or under-exposed. There are lots of techniques available for face detection. Here we used Image segmentation and image filling methods to detect the faces. The next step is Feature extraction which is preprocessing level for age estimation and gender verification. Feature extraction is specific form of dimensionality reduction in image processing and pattern recognition. we use an eigen face using PCA (Principal Component Analysis) method [5] to recognize the faces from the database. Eigen faces are group of Eigen vectors mainly used in computer vision. This approach was developed by Sirovich and Kirby. We used eigenface mainly for age estimation.For Gender classification we used Fisherface algorithm. It uses Principal of Linear Discriminant Analysis method. Using this we find out whether the given image belongs to the gender female or male. The efficiency of the algorithms are checked using K-Cross fold validation and Leave one out classifiers. The classifiers when used over the few publicly available databases, gives out prominence results. The design of the system is showed in the figure 1. It has two main parts namely Enrolment and verification.

METHODS AND TECHNIQUES Histogram Equalization
Histogram equalization is method of transforming contrast. This method is efficient because, it uses intensity values alone. We used Brightness Preserving Dynamic Fuzzy Histogram Equalization (BPDFHE) to improve its brightness and contrast enhancement abilities while decreasing its complexity of computation. This method uses fuzzy statistics of images for their processing and representation. In the fuzzy domain representation and processing of images allows the technique to handle the inaccuracy of gray level values in a better way, resulting in improved performance. Moreover, the inexactness in gray levels is handled well by fuzzy histogram, fuzzy statistics, when calculated with suitable fuzzy membership function, does not have random oscillations or missing intensity levels and is basically smooth. This helps in attaining its significant partitioning needed for brightness preserving equalization. The BPDFHE technique consists Fuzzy Histogram Computation, Partitioning of the Histogram, Dynamic Histogram Equalization of the Partitions, Normalization of the image brightness.

Face Detection
For Face Detection we used Techniques like Image segmentation and image filling to detect the faces. Initially the image is segmented by subtracting the image with the RGB values of the skin tone color. Then the image is turned to pure black where ever the color pixels are found after the subtraction of the skin tone. Then we get the image of faces alone. The noises in the image are filled with the white pixels by calculating the surrounding spaces intensity. If the surrounding pixels are of white then the center pixels are changed to white again. This should be done on a single loop alone. At the last remaining low space dimensional areas of white pixels are removed used the image filling method. After the face regions are separated, the faces are surrounded with a bounded box with the face as a center and the image is cropped. This process is a pre-process for Eigenface recognition. These images are made into templates and sent as input for the Eigen Face recognition. Next step, determine which face class gives the best description for the given input image. This is through minimizing the Euclidean distance

FEATURE EXTRACTION Eigenface Algorithm
If εk is bellow an established threshold θε, the input face is consider to belong to a class. Then the input face image is taken into account of a known face. If the difference is more than given threshold, but bellow a second threshold value, the image can be considered as a unknown face. If the input image is above those two threshold values, the image is determined NOT to be a face image.

Comparison of age stages
In this step we use comparison algorithm to compare the test image with the reference image set stored in the database and the matching score is given as output. For the comparison, we need a database of images with their respective ages specified. Then each image is compared with the input images and the most common features available is selected and the age is determined.

Fisherface Algorithm
It designed using Fisher's Linear Discriminant Analysis which maximizes the ratio of between-classes to within-classes scatter. The plan is simple: same classes should cluster together tightly, while different classes are as far away as possible from each other. The algorithm is described as follows: Construct the Image matrix X with each column representing an image. Each image is a allocateed to a class in the corresponding class vector C. Project X into the (N-c)-dimensional subspace as P with the rotation matrix WPca recognized by a Principal Component Analysis, where N is the number of samples in X and c is unique number of classes.
Find the between-classes scatter of the projection P as, Where mean is the total mean of P, mean_i is the mean of class i in P and N_i is the number of samples for class i.
Calculate the within-classes scatter of P as Where X_i are the samples of class i, x_k is a sample of X_i, mean_i is the mean of class i in P.
Apply a standard Linear Discriminant Analysis and maximize the ratio of the determinant of between-class scatter and within-class scatter. The result is given by the set of simplified eigenvectors Wfld of Sb and Sw corresponding to their eigenvalue. The rank of Sb is atmost (c-1), so there are only (c-1) non-zero eigenvalues, cut off the rest.Finally obtain the Fisherfaces by

Gender classification
After obtaining features via fisherface method, features of the training dataset will be compared with the features of input image to classify whether the given input image is female or male gender.

RESULTS
The experimental results demonstrate that the accuracy of the prediction increases when the dataset provided for the training increases. The algorithms we have discussed so far have been implemented using the MATLAB and has been tested with many different possible datasets which are available publicly in the internet. These datasets for images vary with different constraints like illumination, face expression and many other things like pose etc., The dataset is trained with different types of datasets because it can be used in any generic application. The implementation is then tested with many possible constraints to check for accuracy and errors. And the results have been really prominent. Eigen Faces: We perform the analysis based on the Leave one class out classifier. The recognition rate was found to be 83.5% and we are using 165 models. And each model needs a computation need of 3 seconds and which leads a total computation of 8 minutes. Error rate and reduced space: Fig6.leave one out result Then we perform the analysis with the K-fold cross validation. The K-cross fold validation is performed with 5 folds on dissimilar no. of components. The recognition rate is as follows.