作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程

• •    

基于局部分离与多尺度融合的图像超分辨率重建

  • 发布日期:2023-11-29

Image super-resolution reconstruction based on partial separation and multiscale fusion

  • Published:2023-11-29

摘要: 目前基于深度学习的超分辨率重建网络存在卷积运算冗余、图像重建信息不完整、模型参数庞大,限制了对边缘设 备的适用性等问题。针对上述问题,提出一种轻量级的局部分离与多尺度融合图像超分辨率重建网络,该网络利用局部卷积 对图像进行特征提取,通过分离部分图像通道,在减少网络冗余计算的同时保持图像重建的质量。同时设计一种多尺度特征 融合模块,在空间维度学习长依赖特征,并采用一个通道注意力增强组在空间维度捕获空间特征,减少图像重建信息的丢失, 有效恢复图像的细节纹理。最后,由于多尺度特征融合模块更多的是从全局角度进行特征提取融合,因此构建一种高效反残 差模块补充网络的局部上下文信息提取能力。在 Set5、Set14、B100、Urban100、Manga109 五个基准数据集上进行测试。尺 度因子为×2、×3、×4 时,网络的参数量分别为 373K、382K、394K,浮点运算分别为 84.0G、38.1G、22.1G。定量和定性 实验结果表明,与 VDSR、IMDN、RFDN、RLFN 等网络相比,在较少网络参数的情况下,保证了图像重建效果。

Abstract: Currently, deep learning-based super-resolution reconstruction networks suffer from issues such as convolution operation redundancy, incomplete image reconstruction information, and large model parameters that limit their applicability to edge devices. To address these issues, a lightweight image super-resolution reconstruction network based on partial separation and multiscale fusion is proposed. This network utilizes partial convolutions for feature extraction and separates partial image channels to reduce redundant computations while maintaining the quality of image reconstruction. At the same time, a multiscale feature fusion module is designed to learn long-range dependencies features in the spatial dimension and capture spatial features in the spatial dimension using a channel attention enhancement group, reducing the loss of image reconstruction information and effectively restoring the details and textures of the image. Finally, due to the multiscale feature fusion block focus on global feature extraction and fusion, an efficient inverted residu- al block is constructed to supplement the ability to extract local contextual information. The network is tested on five benchmark da- tasets: Set5, Set14, B100, Urban100, and Manga109, with scale factors of ×2, ×3, and ×4. The parameters of the network are 373K,382K, and 394K, and the floating-point operations are 84.0G, 38.1G, and 22.1G, respectively. Quantitative and qualitative experi- mental results show that compared with networks such as VDSR, IMDN, RFDN, and RLFN, the proposed network ensures image reconstruction quality with fewer network parameters.