Data augmentation guided Decouple Knowledge Distillation for low-resolution fine-grained image classification

Haotian Zhang, YongLiang Qiao, Meili Wang

Primary Area: representation learning for computer vision, audio, language, and other modalities

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Data augmentation/Knowledge Distillation/low-resolution/fine-grained image classification

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: The method proposes a viable solution for fine-grained classification at low resolution.

Abstract:

Continuous development of convolutional neural networks has shown good performance for fine-grained image classification by identifying fine features in high-resolution images.However, in the real world, many images are due to camera or environmental restrictions. Low resolution images with fewer fine features result in a dramatic reduction in classification accuracy.In this study, a twophase Data Augmentation guided Decoupled Knowledge Distillation (DADKD) framework is proposed to improve classification accuracy for low-resolution images.In the proposed DADKD, one phase is data augmentation that generates a composite image and corresponding labels. Another stage is knowledge distillation, which minimizes differences between high-resolution and low-resolution image features.
The proposed DADKD validated on three fine-grained datasets (i.e Stanford-Cars, FGVC-Aircraft, and CUB-200-2011 datasets).
Experimental results show that our proposed DADKD achieves 88.19%, 78.98% and 80.33% classification accuracy on these three datasets, state-of-the-art methods such as SnapMix and Decoupled Knowledge Distillation (DKD).The method proposes a viable solution for fine-grained classification at low resolution.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 5521