A Review of Frequency-Domain Transformers in Image Processing Applications

As one of the mainstream models in image processing, the Transformer has achieved significant breakthroughs in relevant practical applications through extensive experimental validation. However, it also exhibits notable shortcomings in image feature representation learning, including local information loss, high parameter count, and substantial computational complexity. In recent years, the frequency-domain Transformer has emerged as a research hotspot within image processing. This architecture leverages the advantages of frequency-domain transformations in feature extraction, compression, and denoising to achieve deep integration with the Transformer. It partially overcomes the weaknesses inherent in traditional Transformers, fully exploiting the strengths of frequency-domain extraction and Transformer modelling, and demonstrates excellent performance across various image processing tasks. This paper provides an in-depth review of frequency-domain Transformers. It first summarises traditional Vision Transformer architectures and image frequency-domain transformation techniques, introducing the concept of frequency-domain Transformers and outlining their advantages. Subsequently, it comprehensively categorises major types of frequency-domain Transformers and representative models, illustrating their applications across diverse image processing domains from multiple perspectives. Finally, it conducts an in-depth analysis of future research directions for frequency-domain Transformers, offering forward-looking insights.