Enhancing backscatter X-ray imaging with CNN-based denoising: Superior performance in noise reduction and processing efficiency

This study proposes a CNN approach for denoising backscatter X-ray images, which are typically raucous due to low photon energy and subject movement. In contrast to the non-local means filtering and K-SVD, which are highly computationally intensive and require elaborate parameter tuning, the proposed CNN approach maximizes denoising quality while minimizing computation time. The tests were carried out using PSNR and SSIM metrics, and the proposed method shows a clear margin above the traditional methods in performance throughout, at processing times of under 2 s. The influence of batch normalization is assessed in this study; results reveal that smaller batch sizes enhance learning efficiency and decrease overall network error. It shows therefore the ability of CNNs to improve images in real-time while maintaining high fidelity for security and inspection applications, thus providing a scalable and efficient solution for real-time deployment in critical environments.