Dual Focus Net based on SwinIR in Image Super Resolution
Qi Cheng, Xian Fu, Zhuzhu Zhang, Yaqiang Cao, Yu Tao Sun, Hui Zhang · 2024
Image super-resolution (SR) is a critical component in enhancing image quality in computer vision. Traditional Convolutional Neural Networks (CNNs), while effective, often face challenges in handling long-range dependencies and content-independent interactions. Recent advancements with Transformer models, especially Swin Transformer, have opened new avenues in image restoration, but they still require further enhancements in pixel utilization and detail capture. In this study, we introduce SwinDFN (Swin Transformer-based Dual Focus Net), a novel network that addresses these challenges. It integrates the strengths of Transformer-based architectures and enhances them with channel attention and depth-wise convolution in the feed-forward network. Channel attention enables the model to activate more pixels to leverage additional information for enhancing model performance. Concurrently, depth-wise convolution focuses the model more on the restoration of high-frequency information in images. The combination of both further augments the model's feature representation. Our experiments across multiple datasets reveal that SwinDFN surpasses existing CNN-based models and exhibits significant improvements over the SwinIR model. SwinDFN's advancements in SR showcase its potential as a robust solution for high-quality image reconstruction, paving the way for future developments in image processing.