Deep Feature Selection And Mapping For Cross-Age Face Retrieval

Benefit from the recent flourish of deep learning,lots of breakthroughs have been made in face recognition,face verification,and face retrieval fields.Some of latest researches even achieve accuracy better than human beings.Most of the existing face databases have already cover pose,expression and illumination variations,therefore,the algorithms based on these databases can successfully eliminate those variations.However,if an application contains images from a huge gap of age,a new kind of variation has to be considered,which is cross-age variation.Among all the face related research,there is only limited number of them take ageinvariance into account.This paper proposes several novel methods in cross-age face retrieval field.Despite the difficulty of this task,we believe there are many applications that might require a good performance on cross-age face retrieval.For example,how to find an escaped prisoner that has been missing for many years,or how to locate a missing child with no recent pictures.Due to the gap of age between gallery set and probe set,most of the existing method will fail,which makes age-invariant face retrieval remains a challenge.In recent years,many researchers start to focus on age-invariant face feature.We conclude all previous methods into three categories.The first category is the generative method,they utilize aging model to map input face feature into all other ages.Most of the early methods use this strategy.Instead of achieving age-invariance,they generate face features of different ages by only one image.However,it enlarges the gallery sets extremely.The second category called reference coding.They use sparse coding and maximum pooling with the help of a reference set to extract age-invariant face feature.And the third one is the discriminative method.Unlike previous two categories,these research don't rely on aging modeling,which makes it more general and robust.The proposed deep feature selection and mapping is a discriminative approach.Since previous state-of-the-art methods on CACD database are reference coding methods.In this paper,we also develop an improved reference coding method.The proposed two kinds of approaches rely on deep feature and handcrafted feature respectively.We will give a thorough discussion about these algorithms in the experiment section.