The Impact of the Percentile-Clipped Phase Channel on Vision Transformer Performance
Year: 2026
Type: conference-paper
Abstract: Deep learning-based image classification models exhibit excessive reliance on texture information rather than object shapes. This tendency has been identified as a major factor degrading generalization performance and robustness. Although Vision Transformer (ViT) is known to exhibit relatively higher shape bias compared to conventional CNN architectures, standard RGB input representations contain mixed texture and structural information, limiting explicit learning of shape-related cues. To address these limitations, this paper employs Fast Fourier Transform (FFT) to extract phase components that contain structural characteristics of images. A four-channel input representation is proposed by integrating extracted phase component with conventional RGB channels. Experimental results on CIFAR-10 dataset demonstrate that proposed four-channel input representation achieves improved classification accuracy compared to standard RGB input. Through ablation study, significant performance degradation is observed when phase channel is removed or perturbed, indicating that ViT effectively utilizes structural cues derived from phase information. This study demonstrates that explicit structural reinforcement in frequency domain can strengthen shape-oriented learning of ViT and consequently contribute to improvements in visual recognition performance.
Institution: Kangnam University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7202264528