基于渐进式特征融合卷积网络的轻量级超分辨重建算法研究

Research on Lightweight Super-resolution Reconstruction Algorithm Based on Progressive Feature Fusion Convolutional Network

超分辨率重建算法大多通过扩展卷积神经网络提取更多特征细节,容易导致计算复杂度的提高和模型参数量的增大。因此,提出一种渐进式特征融合卷积网络的轻量级超分辨率算法,主要以渐进方式聚合多尺度特征,利用多尺度像素注意力机制构建出简洁高效的上采样模块,保证网络效率和模型设计的轻量级别。在此基础上,还提出基于余弦退火学习的训练策略,在不改变模型结构的情况下提高复原图像的质量。

Most super-resolution reconstruction algorithms extract more feature details through extended convolutional neural networks,which can easily lead to an increase in computational complexity and model parameter quantity.Therefore,this article proposes a lightweight super-resolution algorithm for progressive feature fusion convolutional networks,which mainly aggregates multi-scale features in a progressive manner and utilizes multi-scale pixel attention mechanism to construct a concise and efficient up-sampling module,ensuring network efficiency and lightweight model design.On this basis,a training strategy based on cosine annealing learning is also proposed to improve the quality of restored images without changing the model structure.