Addressing challenges in real-world image classification : long-tailed distribution and knowledge distillation

In computer vision, image classification has progressed rapidly with deep learning over the ten years. However, in the real world, we still face challenges to apply them when the datasets are highly imbalanced, or in some situations to deploy large networks. From the data perspective, in this th...

全面介紹

Saved in:
書目詳細資料
主要作者: Wang, Yiming
其他作者: Lin Guosheng
格式: Thesis-Master by Research
語言:English
出版: Nanyang Technological University 2022
主題:
在線閱讀:https://hdl.handle.net/10356/155131
標簽: 添加標簽
沒有標簽, 成為第一個標記此記錄!
機構: Nanyang Technological University
語言: English
實物特徵
總結:In computer vision, image classification has progressed rapidly with deep learning over the ten years. However, in the real world, we still face challenges to apply them when the datasets are highly imbalanced, or in some situations to deploy large networks. From the data perspective, in this thesis, we aim to improve data augmentations for long-tailed image classification, where only a few semantic classes possess many samples while most other classes have only a few samples. We propose a novel Hybrid Mixup strategy to increase the sample amount and diversity, where we uncover the efficacy of mixup in the latent space of StyleGAN2. Compared with the traditional mixup method on real images, the mixup images generated from the interpolated latent codes have better quality. Experiments on CIFAR-10-LT, CIFAR-100-LT demonstrate that our proposed Hybrid Mixup consistently boosts the head-, medium- and tail-class classification accuracy compared with the traditional mixup method on real images only. Moreover, our results are on par with the state of the arts or even surpass them in some settings. From the model viewpoint, we particularly research the knowledge distillation, which leverages large models to distill enriched knowledge into smaller ones. Here we focus on the scenario where teachers output one-hot predictions only. We find it still possible for students to boost classification accuracy by directly learning from these one-hot predictions. We further propose Patched One-hot Distillation that models empirical probability for teachers to capture the inter-class relationship. Experiments on CIFAR-100 and ImageNet datasets demonstrate that our proposed method helps students learn better than the baseline that directly learns from both the ground-truth labels and the predictions from teachers.