结合注意力与轴向卷积的荧光图像超分辨率方法-《计算机技术与发展》

文章信息/Info

Title:: A Super-resolution Method of Fluorescence Images Combining Attention and Axial Convolution

Author(s):: CHEN Bo-lun; ZHOU Hang; WANG Tie-jun; YANG Hao; School of Computer Science,Chengdu University of Information Technology,Chengdu 610225,China

Keywords:: cycle-consistent adversarial generative network; axial convolution; attention mechanism; image generation; Unet

摘要:: 针对当前荧光显微图像超分辨率重建方法在非配对数据集上训练难度大及计算机资源消耗过高的问题,提出了一种新的解决方案———AxialAttention-SRNet,巧妙地融合了注意力机制与 3D 轴向卷积技术,旨在优化荧光显微图像的超分辨率处理。 AxialAttention-SRNet 的核心在于其超分辨网络模型,该模型以 CycleGAN 为主体,通过整合 3D 轴向卷积模块与 Unet 网络,有效捕捉图像轴向特征的同时,极大地降低了计算量。此外,为了进一步提升图像重建的质量,在高分辨率鉴别器中引入了注意力模块,显著增强了鉴别器的判别能力,从而帮助模型从更多有效信息中学习。在 Thy1-GFPM 和 MTC-EXM 等数据集上的实验充分验证了 AxialAttention-SRNet 的卓越性能。该方法不仅生成的图像质量优异,而且模型复杂度相对较低,有效减轻了计算负担。更重要的是,AxialAttention-SRNet 在真实数据上也表现出色,为荧光显微图像的超分辨率重建提供了一种高效且实用的新方法。

Abstract:: Addressing the challenges of training difficulty on unpaired datasets and excessive computer resource consumption in current fluorescence microscope image super - resolution reconstruction methods, a novel solution, AxialAttention - SRNet, is proposed. This method ingeniously integrates the attention mechanism with 3D axial convolution technology,aiming to optimize the super-resolution processing of fluorescence microscope images. The core of AxialAttention-SRNet lies in its super-resolution network model,which is based on CycleGAN. By integrating 3D axial convolution modules with the Unet network,it effectively captures axial features of images while significantly reducing computational load. Furthermore,to further enhance the quality of image reconstruction,an attention module is introduced into the high -resolution discriminator,significantly boosting its discriminative ability and thereby assisting the model in learning from more informative data. Experiments conducted on datasets such as Thy1-GFPM and MTC-EXM fully demonstrate the out-standing performance of AxialAttention-SRNet. The proposed method not only generates images with superior quality but also features relatively low model complexity, effectively alleviating computational burdens. More importantly, AxialAttention - SRNet exhibits excellent performance on real data,providing an efficient and practical new approach for super-resolution reconstruction of fluorescence microscope images.