A new method to estimate ages of facial image for large database

As a common consensus that the appearances of different persons with the same age diverge widely, we have an opinion that the estimated result of a facial image should be a dynamic range or discrete candidate ages, not a specific one or classified into predefined age groups. Therefore, this paper presents a new method to estimate a set of possible ages of a facial image for large image database with a novel measurement. Firstly by transferring the shape and appearance features of a face into a set of Landmark-Terms, then the famous technology TFIDF in information retrieval and text mining fields is introduced to build a weight matrix of these Landmark-Terms for all age groups, and then some possible ages of a facial image are estimated by this matrix. Secondly, a new clustering method is also used to find the density peaks for each age group by processing the LBP features, then according to the distances of the facial image to the peaks, we obtain another possible estimated ages. Thirdly, we find the first density peak among the two sets of possible ages mentioned above, then choose those ages whose distances to the peak age are short enough in the two set as final estimated ages. Finally, a novel measurement is proposed to evaluate the performance for methods that provide more than one possible estimated ages. The experiments show that our method is promising, the best MAE and CS are close to the best performance of state-of-the-art, and the best BPMAE and NBPMAE also indicate the top possible ages could cover the neighborhood of the the ground-truth age with small errors, in other words, it narrows the age scope effectively.