2D-DCT and Quantization Accelerator for video codecs on MPSoC FPGA using OpenCL framework and Neutrosophy

Sumalatha S.^1,*, Rajeswari²

¹Research Scholar, Acharya Institute of Technology, Visvesvaraya Technological University, India

²Professor and Head, Acharya Institute of Technology, Visvesvaraya Technological University, India

Emails: sumalatha.disha@gmail.com; rajeswari@acharya.ac.in

Abstract

Video codecs based on lossy compression techniques take advantage of removing redundant data in spatial and frequency domains. The various modes of intra- and inter-predictions help to reduce the redundant information in the spatial domain in standard video codecs like AVC, HEVC, and VVC. Further, the removal of redundant information in the frequency domain is achieved by adaptive quantization of transformed frames obtained after DCT-II or DST transformation techniques. In traditional video codec standards, adaptive quantization matrices are derived using the Human Visual System (HVS) model and display resolution parameters, which adjust the quantization step size to preserve perceptually significant pixel information in transformed blocks. The Neutrosophic (NS)-based approach introduces a more refined mechanism for generating the quantization matrix, utilizing Neutrosophic set membership values (true, indeterminate, and false) assigned to each region or frequency component of the transformed block. These values reflect the certainty of pixel relevance, enabling a more adaptive, perceptually driven quantization process. The proposed method incorporates NS logic in combination with the Human Visual System (HVS) model and display resolution parameters. By blending these factors, the quantization step size is optimally tuned to enhance visual quality. The HLS implementation of the transformation and quantization technique suitable for video codec acceleration using the OpenCL framework is adopted in our work. The design was implemented and tested on the Xilinx ZCU-104 board using a standard test sequence from the JCTVC and UVG datasets of various resolutions and diversified content. The testing showed an optimized resource utilization of 60.36%, with notable metrics indicating perceptually good results. The objective metrics showed an improvement of 3.77% in PSNR and 1.83% in SSIM compared to standard HVS-based quantization.

Keywords: Discrete Cosine Transformation; Human Visual System; Vitis HLS; Neutrosopic matrices; Adaptive Quantization; Structural Similarity Index Metric