Background After median filtering, the noise in flash x-ray radiographs is predominantly gaussian. This noise participates in subsequent operations, such as density inversion and interface position extraction, severely compromising the accuracy of density reconstruction and object boundary extraction. Therefore, gaussian noise removal constitutes a critical focus in flash x-ray radiographs denoising research.

Purpose Many methods are available to achieve some degree of denoising, but to meet the high-precision requirements of quantitative analysis in flash x-ray radiographs processing, it is necessary to further reduce the residual noise in flash x-ray radiographs.

Methods Given the small number of flash x-ray radiographs and the presence of mixed noise in flash x-ray radiograph, after investigating neural network denoising methods and studying the U-Net architecture, this paper propose a modified U-Net architecture tailored for flash x-ray radiographs denoising: to prevent excessive noise signals from being introduced through the decoding path, skip connections are only used in the lower layers; to prevent negative values, the sigmoid activation function is used throughout the network to provide nonlinear representation capability.

Results The effectiveness of this network architecture was evaluated through both numerical simulation image and real experimental image, numerical simulation image achieve lower root mean square error (RMSE) and higher signal-to-noise ratio (SNR) and peak signal-to-noise ratio (PSNR), while the profile lines of real experimental image become smoother after denoising.

Conclusions Experiment results demonstrate that this method achieves significantly improved denoising performance. By adjusting the network structure, better results may be achieved in the future.