ML-Driven Denoising Framework for Structured Light Modes in Realistic Free Space Channels

This work compares a convolutional autoencoder (CAE) enhanced with morphological feature extraction against the classical BM3D algorithm for denoising structured light modes. Using Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) for evaluation, results show that while BM3D yields competitive PSNR, the proposed approach consistently achieves higher SSIM, preserving structural details more effectively. At the highest noise level 5 dB SNR, the proposed method achieved an average PSNR of ∼ 41 dB and SSIM above 0.994, significantly outperforming BM3D, which reached 39.2 dB PSNR and 0.92 SSIM. Our approach not only improves noise suppression but also preserves essential structural features, making it a robust solution for optical mode restoration under severe noise.