Enhanced age estimation by considering the areas of non-skin and the non-uniform illumination of visible light camera sensor

Most previous research on human age estimation based on the detection of multiple feature points using the active appearance model (AAM) method. However, it is difficult to use the AAM-based methods in actual applications, because their performance is strongly affected by image backgrounds, head movements, and non-uniform facial region illumination. Furthermore, they require significant processing time. Other age estimation methods based on a detected face box area may be considered as an alternative; however, noise areas that include hair, backgrounds, and non-uniform illumination of visible light camera sensor may be inadvertently included in the face box, which reduces age estimation accuracy. Therefore, we propose a new age estimation method that is robust to these noise areas. Our proposed method is novel in following four ways. First, we propose an age estimation method using a weighted multi-level local binary pattern (wMLBP) based on a fuzzy-logic system. Second, two input values (the difference between the mean gray levels of the sub-block and the central area of the face, and the distance from the sub-block to the center of the facial region) are determined considering the noise areas of hair, background, and non-uniform illumination of visible light camera sensor. Then, the optimal weights are determined using a fuzzy-logic system with these two input values, which does not require a time-consuming training process. Third, by assigning an optimal weight to the histogram features extracted by the MLBP method in each sub-block, age estimation accuracy is enhanced. Finally, the age is estimated using a SVR method based on a combination of weighted MLBP features and Gabor wavelet features. Experimental results obtained using the public PAL and MORPH age databases demonstrate that the accuracy of our method is superior to other previous methods.