Deep feature augmentation for occluded image classification

Due to the difficulty in acquiring massive task-specific occluded images, classification of images with deep convolutional neural networks (CNNs) remains highly challenging. To alleviate dependency on large-scale image datasets, we propose a novel approach improve accuracy by fine-tuning pre-trained models set augmented feature vectors (DFVs). The DFVs is composed original and pseudo-DFVs. pseudo-DFVs are generated randomly adding difference (DVs), extracted from small clean pairs, real DFVs. In fine-tuning, back-propagation conducted DFV data flow update network parameters. experiments various datasets structures show that augmentation significantly improves without noticeable influence performance images. Specifically, ILSVRC2012 dataset synthetic proposed achieves 11.21% 9.14% average increases for ResNet50 fine-tuned occlusion-exclusive occlusion-inclusive training sets, respectively.