Merged Multi-CNN with Parameter Reduction for Face Attribute Estimation

This paper proposes a face attribute estimation method using Merged Multi-CNN (MM-CNN). The proposed method merges single-task CNNs into one CNN by adding merging points and reduces the number of parameters by removing the fully-connected layers. We also propose a new idea of reducing parameters of CNN called Convolutionalization for Parameter Reduction (CPR), which estimates attributes using only convolution layers, in other words, does not need any fully-connected layers to estimate attributes from extracted features. Through a set of experiments using the Celeb A and LFW-a datasets, we demonstrated that MM- CNN with CPR exhibits higher efficiency of face attribute estimation than conventional methods.