Hidformer: A Frequency-Domain Transformer for High-Resolution Image Deblurring in IoT Devices

Image deblurring is a critical technology for enhancing the visual perception capabilities of Internet of Things (IoT) devices, but challenges such as motion blur and defocus blur continue to degrade image quality. Although convolutional neural networks (CNNs) have achieved progress in image deblurring tasks, their limited receptive fields and static filters result in performance bottlenecks when processing high-resolution images and modeling long-range dependencies. Transformers, while demonstrating remarkable success in vision tasks, suffer from computational complexity that increases quadratically with spatial resolution, making them less suitable for resource-constrained IoT environments. To address these challenges, we propose Hidformer, a novel architecture that combines frequency-domain processing with Transformer structures to efficiently capture multiscale representations and model long-range dependencies, particularly for high-resolution image deblurring. Inspired by the convolution theorem, we introduce the frequency-domain self-attention (FDSA) module, which leverages fast fourier transform (FFT) to map images from the spatial-domain to the frequency-domain and computes attention maps efficiently through the Hadamard product, significantly reducing spatial and temporal complexity. To further optimize performance, we propose SIGA, a simplified variant of the gated linear unit (GLU) that replaces the computationally expensive gaussian error linear unit (GELU) while retaining nonlinearity. Additionally, we design an optimal cutoff-frequency selection algorithm (OCSA) to improve the performance of high-pass filters. Extensive experiments on GoPro dataset demonstrate that Hidformer outperforms state-of-the-art methods, especially in detail and texture recovery. With only 16.3 M parameters and an inference time of 0.11 s, Hidformer achieves a balance between accuracy and efficiency, highlighting its potential for deployment in IoT devices, particularly in edge computing environments.