Frequency-Split Inception Transformer for Image Super-Resolution

Abstract

Transformer models have shown remarkable effectiveness in capturing long-range dependencies and extracting features for single image super-resolution. However, their deployment on edge devices is hindered by their high computational complexity. To address this challenge, we propose Inception Swin Transformer (IST), a novel model that leverages frequency domain separation to reduce redundant computations.In IST, we exploit the strengths of both CNN-based networks and Transformer variants to handle high-frequency and low-frequency features, respectively. By dynamically utilizing frequency factors to separate feature maps, IST ensures that different components are processed appropriately. Additionally, IST maintains a balanced trade-off between model speed and performance by gradually reducing the proportion of high-frequency components.Our experiments demonstrate that IST effectively reduces the FLOPs while preserving high performance. The combination of Transformers’ accuracy and CNN variants’ efficiency enables IST to significantly reduce computational strain without compromising quality. Comparative analysis reveals that IST outperforms other models, achieving superior results with less FLOPs.