Local Phase Quantization Feature Extraction based Age and Gender Estimation Using Convolutional Neural Network

Date

2017

Journal Title

Journal ISSN

Volume Title

Publisher

Tartu Ülikool

Abstract

Even though artificial neural networks are one of the oldest machine learning techniques, there were no many experiments on them by 2010s because of its computational complexity. Artificial neural networks got inspired by human neural anatomy, and try to achieve similar accuracy. Latest advances of silicon technology enable us to conduct experiments on all types of artificial neural networks. Convolutional Neural Networks are one of state-of-art neural network types. As a human, we all have great recognition, detection mechanism in our body. In this study, it will be attempted to gain similar ability with computer aid of CNNs. As all other supervisedlearning methods, we need training and testing dataset. We are going to apply CNN on apparent age and gender estimation. There are few public dataset which are created for age estimation. One of them and the biggest one is IMDB-Wiki dataset which contains pictures of famous people from wikipedia and IMDB with their real-age label. In order to create real-age label, the creator used the time differences between photo-taken year and birth year. However for better accuracy, we need apparent age information. Because aging is a process that depends on many conditions. As it is going to be explained later, we collected Japanese dataset on the internet, and labeled their apparent ages by weighted voting. After collecting the image data sets, we pre-processed the images with face detection and alignment methods. Afterwards, we copied all images and used Local Phase Quantization(LPQ) method to extract their features. In CNN, it is always better to use pre-trained data and fine-tune it. Thus we used deep face recognition pre-trained data with almost 2 millions images. After that, we fine tuned images(with LPQ and without LPQ separately) with using the label distribution encoding. Finally we had 2 CNN data. For combining the results, we took the mean of all respective output neurons. At the end, expected values of all neurons are considered as apparent age information. For gender classification, we just trained the system in the similar way. Only difference is that we have only 2 output neurons for gender classification, besides LPQ is not applied in gender classification.

Description

Keywords

histogram equalization, face detection, face alignment, label distributed encoding, local phase quantization, neural network, convolutional neural network, deep learning, VGG- 16, VGG-19, age estimation, gender classification

Citation