基于USENet实现数字全息细胞再现相位像超分辨重构

在UNet框架中集成SENet权重标定学习机制,设计了USENet实现图像超分辨重构.网络以对称拓扑叠加残差运算,通过多尺度卷积窗口增强模型精度与泛化能力,为图像特征通道引入权重标定算法以提升兴趣区域的估算置信度.结果表明,该模型能够明显改善输入样本的细节特征,已将验证集与真实图像的结构相似性从平均0.770 2提升至0.942 7.根据实验组和不带标定层的对照组重建图像对比显示,标定层可进一步在全局图像中将兴趣区域从平均0.965 5提升至0.970 3.

Super-resolution in Digital Holographic Phase Cell Image Based on USENet

This research integrates learning mechanisms of weights calibration and multiple receptive fields in SENet into UNet,working out USENet to achieve super-resolution in digital holographic phase cell images.Structured with symmetrical topology and filtered with multi-scale,the model is trained to improve the image rebuilding accuracy and the capability of generalization.So as to enhance the estimation confidence in the region of interest,the weights calibration blocks are introduced to differentiate the importance of feature map channels.With better visual effects and details,the experimental results confirm that the average numerical score of structural similarity index on validation set has been verified to improve from 0.7702 to 0.9427.According to the performances between experimental group of USENet and reference group of network without calibration layer,the region of interests in global images have demonstrated a further improvement from 0.9655 to 0.9703 by the block of calibration.