Joint feature extraction and classification in a unified framework for cost-sensitive face recognition

Highlights：

• In this paper, we concern the cost-sensitive application of the door-locker system based on face recognition, which contains the high-dimensional and cost-sensitive problems.

• We propose a unified cost-sensitive learning framework for the door-locker system based on face recognition. The proposed framework can optimize the learning of feature representations and the classifier jointly in an iterative manner. This enables cost-sensitive processes of feature extraction and classification that can update the feature representations and classifier information simultaneously with the overall misclassification loss of all training images minimized.

• We propose to learn the face features by preserving the cost-sensitive local manifold structure. In this way, we introduce a cost-sensitive regularization to guide the feature extraction process, which can make the learned feature representations discriminative and cost-sensitive.

• We evaluate the proposed method with 11 state-of-the-art face learners and 9 cost-sensitive methods. The experimental results on three public face benchmarks demonstrate that the proposed method can significantly reduce the overall misclassification loss of face recognition system as well as the classification errors associated with high costs. Especially when comparing with the state-of-the-art face learners, the proposed method can achieve up to 58.51% total cost reduction on CMU PIE dataset and 90.64% improvement for imposter detection on LFW-a dataset.