Evaluation of CNN- and COSFIRE-based age group classification from face images

Abstract

In this project, we focus on the age group classification problem where a face image is classified into one of the predetermined age groups. Age prediction has many potential applications such as data search optimization, personalized human-computer interaction, commercial profiling, access control and surveillance. As evidence suggests, humans themselves are not quite accurate in age estimation and therefore, the development of an accurate automatic age estimation system is very valuable and intriguing [1]. However, age prediction comes with various challenges such as low image quality, lack of labeled benchmark datasets and modern cosmetic treatments. Though, the main difficulty is the fact that the aging process differs greatly from person to person.

In this work, age group classification is performed using two methods; a pretrained CNN that uses a deep learning approach called VGG-Face and a trainable filter approach named COSFIRE. COSFIRE has shown very good results in many different applications [2{7] and it is thus used in this project to analyze its performance in further areas. On the other hand, being that recent literature suggests that CNNs deliver the highest accuracy rates within such problems, a CNN was chosen to compare COSFIRE's results with. Being one of the current best performing CNNs for this task, VGG-Face was chosen as our CNN.

Using both methods, feature vectors are extracted from the labeled FERET Dataset images and utilized for the training of the Support Vector Machine (SVM) classifiers to construct models that classify the images correctly. Face detection and rescaling is applied to the images prior to the feature extraction to remove the unnecessary background. Boot-strap aggregatating was applied to avoid biased learning due to the unbalanced data but results show that the system performed better without such method. Further experiments have shown that the best result is achieved when using the penultimate layer of the pre-trained VGG-Face as the output layer and when configuring 90 filters for COSFIRE. The proposed combined method, using the feature vectors extracted from both VGG-Face and COSFIRE to train a set of stacked SVMs, has shown its applicability to the age classification task by obtaining an accuracy of 97.8%, even when used images have illumination, pose and expression variations.