Deep feature augmentation for occluded image classification
作者：
Highlights：
• The paper proposes a deep feature vector augmentation approach for end-to-end learning to boost the classification accuracy for occluded images.
• The proposed approach requires only a small set of clean and occluded image pairs and thus is suited for practical applications.
• The model fine-tuned with the proposed approach does not require occlusion detection in inference and has an almost unnoticeable influence on the classification accuracy for clean images. Therefore, it is universal for the classification of both the occluded images and the clean images.
• A large-scale occluded image dataset is synthesized from ILSVRC2012 dataset.
• Significant improvement on occluded image classification is demonstrated on various real and synthetic occluded image datasets.
摘要
•The paper proposes a deep feature vector augmentation approach for end-to-end learning to boost the classification accuracy for occluded images.•The proposed approach requires only a small set of clean and occluded image pairs and thus is suited for practical applications.•The model fine-tuned with the proposed approach does not require occlusion detection in inference and has an almost unnoticeable influence on the classification accuracy for clean images. Therefore, it is universal for the classification of both the occluded images and the clean images.•A large-scale occluded image dataset is synthesized from ILSVRC2012 dataset.•Significant improvement on occluded image classification is demonstrated on various real and synthetic occluded image datasets.
论文关键词：Deep feature augmentation,Image occlusion,Convolutional neural networks,Image classification
论文评审过程：Received 24 March 2020, Revised 19 September 2020, Accepted 29 October 2020, Available online 31 October 2020, Version of Record 5 November 2020.
论文官网地址：https://doi.org/10.1016/j.patcog.2020.107737
