结合整体注意力与分形稠密特征的图像超分辨率重建

现有单图像超分辨率模型普遍基于卷积神经网络且使用单一尺度的卷积核提取特征信息，容易造成细节信息遗漏并降低网络表征能力。为有效提取高频信息同时提高图像重建性能，提出一种基于整体注意力机制与分形稠密特征增强的图像超分辨率重建模型。在特征增强过程中，级联9个分形稠密特征增强模块，每个模块通过4条分支路径提取和融合多尺度特征，并引入局部稠密跳跃连接传递信息以获取更丰富的细节信息。引入整体注意力机制，从3个维度出发建立特征图之间的关联关系，通过对不同通道、空间和层次的特征进行加权和选择性聚合为特征图分配不同的权重，从而提高模型判别学习能力。在Set5、Set14、BSDS100和Urban100数据集上的实验结果表明，该模型可有效重建纹理细节更丰富的高分辨率图像，重建图像在主观视觉效果与客观评价指标上均优于同类模型，且在图像放大3倍时，峰值信噪比和结构相似性指标最高比MSRN模型提升了0.57 dB和0.007。

Image Super-Resolution Reconstruction Combining Holistic Attention and Fractal Density Feature

Most current single-image super-resolution models are based on Convolutional Neural Networks (CNN) that use a single-scale convolution kernel to extract feature information, resulting in the missing details and reduced network representation ability.This study proposes an image super-resolution reconstruction model based on a holistic attention mechanism and fractal density feature enhancement to effectively extract high-frequency information and improve reconstruction performance. During the feature enhancement process, the proposed model cascades nine Fractal Dense Feature Enhancement(FDFE) modules.Each module extracts and fuses multiscale features through four branch paths and introduces local dense jump connections to transfer information for enhanced, detailed information.Next, the proposed model uses a holistic attention mechanism to establish the correlation between feature maps in three dimensions and assigns different weights to feature maps by weighting and selectively aggregating features from different channels, spaces, and layer to improve the discrimination learning ability of the model.The experimental results on the Set5, Set14, BSDS100, and Urban100 datasets show that the proposed model can effectively reconstruct high-resolution images with richer texture details.It is superior to many similar models in terms of subjective visual effects and objective evaluation indicators.When the image is magnified three times, the Peak Signal-to-Noise Ratio(PSNR) and Structural Similarity(SSIM) indexes of the proposed model are up to 0.57 dB and 0.007 higher than the Multi-Scale Residual Network(MSRN) model.