首页|期刊导航|电子科技|基于特征调制蒸馏网络的图像超分辨率重建

基于特征调制蒸馏网络的图像超分辨率重建OA

Image Super-Resolution Reconstruction Based on Feature Modulation Distillation Network

中文摘要英文摘要

针对图像超分辨率算法因模型复杂度较高而难以应用于低性能设备的问题,文中设计了轻量化的特征调制蒸馏网络.提出特征调制提纯模块,通过双分支聚合对特征进行高效提纯,简化模型结构.改进混合注意力模块在空间和通道维度上进一步增强了特征表达,轻量化双向渐进上采样模块通过水平方向和垂直方向的双阶段上采样有效降低了特征重建时的模型复杂度.实验结果表明,所提算法的性能优于现有方法,在模型复杂度和重建性能之间实现了良好平衡.在5个基准测试集的×2超分辨率任务中,相较于信息蒸馏网络,所提算法的峰值信噪比平均提高了约0.45 dB,参数量和乘加累积运算量仅为其44%和41%.

In view of the problem that image super-resolution algorithms are difficult to be applied to low-per-formance devices due to their high model complexity,a lightweight feature modulation distillation network is de-signed.A feature modulation purification module is proposed,which efficiently purifies features through dual-branch aggregation to simplify the model structure.The improved hybrid attention module further enhances feature representation in both spatial and channel dimensions.The lightweight bidirectional progressive upsampling module effectively reduces the model complexity during feature reconstruction through dual-stage upsampling in both hori-zontal and vertical directions.The experimental results show that the performance of the proposed algorithm is supe-rior to the existing methods,achieving a good balance between model complexity and reconstruction performance.In the×2 super-resolution tasks of five benchmark test sets,compared with the information distillation network,the peak signal-to-noise ratio of the proposed algorithm is increased by an average of approximately 0.45 dB,while the number of parameters and the multiply-add cumulative computational complexity are only 44%and 41%respective-ly.

陈鹏民;陈清江

西安建筑科技大学 理学院,陕西 西安 710055西安建筑科技大学 理学院,陕西 西安 710055

信息技术与安全科学

卷积神经网络高效图像超分辨率特征调制信息蒸馏轻量化上采样注意力机制

convolutional neural networkefficient image super-resolutionfeature modulationinformation distillationlightweight upsamplingattention mechanism

《电子科技》 2026 (10)

52-62,11

国家自然科学基金(12101482) National Natural Science Foundation of China(12101482)

10.16180/j.cnki.issn1007-7820.2026.10.006

评论