基于卷积注意力机制的无透镜对抗编码成像

Lens-free adversarial encoded imaging method based on convolutional attention mechanism

相较于单孔径光学无透镜成像技术,编码孔径成像(Coded Aperture Imaging,CAI)技术具有更高的光照强度和更高的分辨率,近年来引起了广泛的关注。然而,现有的CAI技术存在成像时间分辨率较低和成像性能较差的限制。深度学习技术由于其强大的复杂特征建模能力,被广泛应用到各类信号处理领域中。本文提出利用深度学习技术解决上述问题,构建基于卷积注意力机制的生成对抗网络模型(Convolutional Attention Mechanism Based-Generative Adversarial Network,CAM-GAN)来适应不同的情境和任务要求,提高CAI效果与稳定性。通过引入卷积注意力机制模块使生成器选择性地聚焦于数据的特定区域,恢复原始图像的细节和结构。在此基础上,以生成对抗网络的形式进行网络的训练,生成更逼真、更高质量的图像。在公开数据集的实验结果表明,与其他方法相比,CAM-GAN在图像质量上表现出色,取得了最高的峰值信噪比值,相较于次优的UNet-GAN算法提高了约0.32,充分证明了深度学习技术在CAI领域中的应用潜力。

Compared with single-aperture optical imaging technologies,coded aperture imaging(CAI)could achieve higher light intensity and resolution,thus attracting extensive attention in recent years.However,existing CAI technologies have limitations in low temporal resolution and inferior imaging performance.Deep learning technologies have been widely applied in various signal processing fields due to their powerful capabilities in modeling complex features.Therefore,we develop a convolutional attention mechanism based generative adversarial network(CAM-GAN)model to utilize deep learning technologies to address the above problems,which could adapt to different scenarios and task requirements to improve the effects and stability of CAI.Convolutional attention mechanism module is introduced in this model so that the generator can selectively focus on specific areas of the data to recover details and structures of the original image.On this basis,the network is trained in the form of generative adversarial networks to generate more realistic and higher quality images.Experimental results on public datasets show that compared with other methods,CAM-GAN performs excellently on image quality and achieves the highest peak signal-to-noise ratio,improving by about 0.32 over the suboptimal UNet-GAN algorithm,which fully demonstrates the application potential of deep learning technologies in the CAI field.