Desenvolvimento de arquiteturas de redes neurais profundas para melhoria de qualidade em vídeos comprimidos

Abstract

Video consumption currently accounts for the majority of Internet data traffic, driven by emerging technologies such as virtual reality, 360° videos, and increasingly higher resolutions like 4K and 8K. To enable the transmission and storage of such content, video coding standards are employed to significantly reduce data volume. However, this compression process introduces visual artifacts that degrade the perceived qual ity for the end user. Post-processing solutions based on deep learning have proven effective in restoring visual quality. Therefore, the main objective of this thesis was the development of solutions based on deep neural network architectures to enhance the quality of compressed videos, considering different coding standards, compres sion scenarios, and quantization levels. Three main approaches were developed. The first, called multi-codec, was trained with videos compressed using multiple codecs (HEVC, VVC, AV1, and VP9) and demonstrated good generalization, with ∆PSNR gains ranging from 0.091 dB to 0.382 dB. The second approach, multi-domain, treated each codec as a separate domain during training, which increased the specialization potential of the network. This solution achieved superior results, with average ∆PSNR gains between 0.228 dB and 0.764 dB, in addition to bit rate reductions of up to 16.50% (BD-Rate) and consistent improvements in BD-PSNR. The final proposal, 3D-STDF, in corporated three-dimensional convolutions to more effectively and integratively capture spatio-temporal correlations between consecutive frames. This architecture achieved the best results, with a ∆PSNR of up to 0.85 dB (3D-STDF-L model) and a ∆SSIM of 0.016 at QP 37, while maintaining strong performance even under severe quantiza tion (e.g., 0.420 dB at QP 47). The model also achieved a BD-Rate of up to -16.50% and a BD-PSNR of 0.709 dB, outperforming other proposed architectures and demon strating a solid balance between performance and computational cost. These results show that the thesis contributed generalizable and effective solutions for compressed video restoration, balancing enhancement performance, generalization capability, and computational efficiency.