Face recognition with a small occluded training set using spatial and statistical pooling

Occlusion is one of the most intractable problems for face recognition. Double-occlusion problem is an extremely challenging case that the occlusion can occur in both of training and test images. Existing robust face recognition approaches against occlusion rely on large-scale training data, which can be expensive or impossible to obtain in many realistic scenarios. In this paper, we aim to address the double-occlusion problem with a limited amount of training data using a unified framework named subclass pooling. A face image is divided into ordered subclasses according to their spatial locations. We propose a fuzzy max-pooling scheme to suppress unreliable local features from occluded regions. The final average-pooling can enhance the robustness by automatically weighting on each subclass. Our method is evaluated on two face recognition benchmarks. Experimental results suggest that our method leads to a remarkable margin of performance gain over the benchmark techniques.