Analysis of the effect of image resolution on automatic face gender and age classification

In previous work, the effect of image resolution for gender classification performance was analyzed using facial images. This work revealed even very low resolution images such as 3x2, carry useful information for gender detection. Inspired by this study, this thesis seeks an answer to the question of "Do low-resolution images carry some useful information for distinguishing between age groups?". For this purpose, four different feature types namely pixel values, Local Binary Patterns, Multi-Block Local Binary Patterns and Histogram of Oriented Gradients were employed in the experiments where each experiment was repeated using ten different image resolutions including, 2x1, 3x2, 6x5, 8x6, 11x9, 16x13, 22x18, 45x35, 90x72, 329x264. Two popular facial datasets, MORPH and FG-NET Aging databases were used for the experiments. Since the large number of images in MOPRH database has negative effects on the duration of experiments, a subset of MORPH database preserving gender and age rates was used. Some preprocessing operations, e.g., face detection, cropping, aligning eye centers, histogram equalization were applied to the images prior to the classification process. K- Nearest Neigbor and Random Forests algorithms were employed separately for age and gender classifications. Age classification experiments were conducted on MORPH database using separate age groups first. The experiments were then repeated with nested age groups to employ more realistic age classification scenarios. Since nested age groups increased the success rate, such groups were used for FG-NET only. As a result, for gender classification, satisfying results were obtained even for low resolution images such as 11x9. Since age classification is harder and more complicated than that of gender, lower age classification success rates were obtained. For age classification, maximum 59.8% success rate for MORPH database and maximum 57.7% success rate for FG-NET database were recorded. K- Nearest Neighbor results are highly affected by the different image of the same persons. Hence, Random Forest results are taken to be more realistic. In terms of feature types, maximum success rates were obtained using pixel values and Histogram of Oriented Gradients.