Recently, Transformer networks have demonstrated outstanding performance in the field of image restoration due to the global receptive field and adaptability to input. However, the quadratic computational complexity of Softmax-attention poses a significant limitation on its extensive application in image restoration tasks, particularly for high-resolution images. To tackle this challenge, we propose a novel variant of the Transformer. This variant leverages the Taylor expansion to approximate the Softmax-attention and utilizes the concept of norm-preserving mapping to approximate the remainder of the first-order Taylor expansion, resulting in a linear computational complexity. Moreover, we introduce a multi-branch architecture featuring multi-scale patch embedding into the proposed Transformer, which has four distinct advantages: 1) various sizes of the receptive field; 2) multi-level semantic information; 3) flexible shapes of the receptive field; 4) accelerated training and inference speed. Hence, the proposed model, named the second version of Taylor formula expansion-based Transformer (for short MB-TaylorFormer V2) has the capability to concurrently process coarse-to-fine features, capture long-distance pixel interactions with limited computational cost, and improve the approximation of the Taylor expansion remainder. Experimental results across diverse image restoration benchmarks demonstrate that MB-TaylorFormer V2 achieves state-of-the-art performance in multiple image restoration tasks, such as image dehazing, deraining, desnowing, motion deblurring, and denoising, with very little computational overhead.

In the Face Super-Resolution (FSR) task, it is important to precisely recover facial textures while maintaining facial contours for realistic high resolution faces. Although several CNN-based FSR methods have achieved great performance, they fail in restoring the facial contours due to the limitation of local convolutions. In contrast, Transformer-based methods which use self-attention as the basic component, are expert in modeling long-range dependencies between image patches. However, learning long-range dependencies often deteriorates facial textures due to the lack of locality. Therefore, a question is naturally raised: how to effectively combine the superiority of CNN and Transformer for better reconstructing faces? To address this issue, we propose an Efficient Latent Style guided Transformer-CNN framework for FSR called ELSFace , which can sufficiently integrate the advantages of CNN and Transformer. The framework consists of a Feature Preparation Stage and a Feature Carving Stage. Basic facial contours and textures are generated in the Feature Preparation Stage, and separately guided by latent styles, so that facial details are better represented in reconstruction. CNN and Transformer streams in the Feature Carving Stage are used to individually restore facial textures and facial contours, respectively in a parallel recursive way. Considering the negligence of high-frequency features when learning the long-range dependencies, we design the High-Frequency Enhancement Block (HFEB) in the Transformer stream. The Sharp Loss is also proposed for better perceptual quality in optimization. Extensive experimental results demonstrate that our ELSFace can achieve the best results among all metrics compared to the state-of-the-art CNN and Transformer-based methods on commonly used datasets and real-world tasks. Meanwhile, our ELSFace method has the least model parameters and running time. The codes are released at https://github.com/FVL2020/ELSFace .