Quantized Compressive Sensing Video Compression Framework for Low Power Applications

Abstract

Compressive sensing is an approach to reduce the acquired signal dimension with a sampling rate lower than the Nyquist rate. This is achieved by incorporating suitable quantization techniques, such as scalar quantization, to compensate for the low sampling rate. However, quantization of small compressive sensing measurements does not perform well in terms of rate-distortion. This paper presents an adaptive variant of differential pulse code modulation coupled with scalar quantization for quantizing and de-quantizing the compressive sensing measurements of videos. The proposed adaptive approach along with scalar quantization allows reduction of the necessary bandwidth and provides a higher prediction gain than the conventional differential pulse code modulation. The performance of the proposed approach is validated by integrating with various compressive sensing reconstruction schemes and compared with the conventional video codec approach. The reconstructed video frames' visual quality using the proposed approach was also tested based on SSIM metric. The experimental result shows that the proposed scheme provides substantial gain over uniform scalar quantization, and scalar quantization with differential pulse code modulation at various compressive sensing measurements bitrates for different video sequences. When compared with conventional video compression schemes (containing side information at the encoder), the proposed Quantized CS approach shows notable improvements in reconstruction quality when compared with conventional video codecs at lower bitrates. In sum, the proposed methodology could contribute to benefit various battery powered application of visual sensor network in terms of low data processing, less power consumption and minimum memory utilization etc.