Lightweight Facial Super-Resolution Reconstruction of Blueprint Separable Residual Networks
Jiahao Pan, Zhihao Ling · 2024
Current deep learning-driven facial super-resolution models tend to adopt more complex designs, leading to escalating computational complexity and memory requirements. Such designs, focused solely on singular reconstruction performance, not only consume substantial computational resources but also struggle to adapt to practical application scenarios. Addressing these challenges, this paper presents a lightweight dual-branch facial super-resolution reconstruction network, founded on a blueprint separable residual structure. The proposed model leverages reparameterizable blueprint separable convolutions to enhance feature aggregation and improve model performance without incurring additional computational costs. By integrating anti-aliasing and Gaussian blur, the model effectively supervises facial boundary prior information with low operational complexity, while fully utilizing the semantic information of facial images to enhance reconstruction results. Additionally, an ultra-low computational cost attention mechanism module captures positional information and channel relationships, leading to exceptional reconstruction performance. Experimental results demonstrate that the model exhibits significant superiority in terms of reconstruction quality, complexity, and robustness on the CelebA-HQ and Helen test datasets.