【摘要】 近年来，深度学习技术推动了图像超分辨率领域的进展。然而，现有的网络计算资源消耗量大，无法在移动端设备上进行应用。针对这一问题，提出了一种基于特征金字塔的轻量化图像超分辨率算法。在设计的特征金字塔提取模块中，采用通道分离和交叉融合的操作，提高特征的表达能力，更全面地捕捉不同特征间的关联。此外，设计了一种双分支的通道注意力机制，引导两个特征自适应性融合，同时引入残差结构让模型训练更加稳定。在Manga109数据集的2倍尺度的实验中，相比于SRCNN算法和IDN算法，提出的方法峰值信噪比（PSNR）分别提高了1.35 dB和0.18 dB，参数量相比IDN减少了43.76%。更多还原

【Abstract】 This paper proposes a lightweight image super-resolution algorithm based on feature pyramid. In the designed feature pyramid extraction module, channel separation and cross-fusion operations are used to improve the expression ability of features and more comprehensively capture the correlation between different features. In addition, a dualbranch channel attention mechanism is designed to guide the adaptive fusion of two features, and a residual structure is introduced to make model training more stable. In the x2-scale experiment of the Manga109 data set, compared with the SRCNN algorithm and the IDN algorithm, the PSNR of the method proposed in this article increased by 1.35 dB and 0.18 dB respectively, and the number of parameters is reduced by 43.76% compared to IDN.更多还原