基于卷积神经网络和Transformer的高效图像超分辨率重建
Efficient Image Super-Resolution Reconstruction Based on Convolutional Neural Networks and Transformer
李邦源 1杨家全 2薛若漪 3张晓宇 3汪航 3孙宏滨4
作者信息
- 1. 云南电网有限责任公司玉溪供电局,云南 玉溪 653100
- 2. 西安交通大学人工智能学院,陕西 西安 710049;云南电网有限责任公司电力科学研究院,云南 昆明 650217
- 3. 西安交通大学微电子学院,陕西 西安 710049
- 4. 西安交通大学人工智能学院,陕西 西安 710049
- 折叠
摘要
深度学习推动了图像超分辨率重建技术的显著进步,但复杂的操作导致计算和内存成本高昂,限制了其实际应用.为此,提出了一种新颖的算法,融合了Transformer和卷积神经网络,同时采用膨胀卷积和深度可分离卷积技术.在五个基准数据集上的实验证明,所提EHN模型能够高效提取超分辨率特征,在更少参数和推理时间下实现与现有方法相当甚至更好的超分辨率效果.特别地,在×2、×3和×4放大倍数下,EHN的推理时间仅为现有网络的18.4%、18.9%和20.3%,这一优势对于处理大量图像的场景至关重要,能够显著减少计算时间和资源消耗,提升整体效率.
Abstract
Deep learning has significantly advanced image super-resolution reconstruction techniques.However,the computational and memory costs associated with complex operations have limited their practical applications.To address this issue,a novel algorithm is proposed,which integrates Transformer and Convolutional Neural Networks(CNNs)while incorporating dilated convolutions and depthwise separable convolutions.Experimental results on five benchmark datasets demonstrate that the proposed EHN model efficiently extracts super-resolution features,achieving comparable or even better super-resolution performance with fewer parameters and inference time compared to existing methods.Specifically,under×2,×3,and×4 magnification factors,the inference time of EHN is merely 18.4%,18.9%,and 20.3%of that of existing networks,respectively.This advantage is crucial for scenarios involving the processing of large volumes of images,significantly reducing computational time and resource consumption,thereby enhancing overall efficiency.
关键词
图像超分辨率/Transformer/卷积神经网络/膨胀卷积/深度可分离卷积Key words
Image super-resolution/Transformer/Convolutional neural networks/Dilated convolution/Depthwise separable convolution引用本文复制引用
基金项目
云南电网科技项目(YNKJXM20220023)
出版年
2024
NETL
NSTL
万方数据