Multi-resolution extreme learning machine-based side information estimation in distributed video coding

Abstract

Context: Encoding of video frames in a traditional video coding architecture involves exhaustive computations due to the motion estimation (ME) task. Hence, it requires a considerable amount of computing aid, battery power, and resource memory. These codecs are not effective and reliable for applications like surveillance systems, wireless sensor networks, wireless camcorders, having scarcity in the availability of resources and computing ability. Therefore, in such scenarios, distributed video coding (DVC) represents a viable solution for power-constrained hand-held devices. DVC empowers the adaptability in distributing the complexity between the encoder and the decoder. Objective: Like any other building block, the decoder driven side information (SI) generation module plays a key role in a DVC codec. The efficacy of a DVC codec firmly relies on the quality of the SI generated at the decoder. SI is considered to be the facsimile of the original Wyner-Ziv (WZ) frame. Hence, the superior the quality of SI, improved is the efficiency of the codec. The primary objective of the present work is to enhance the quality of the SI frame so that the overall performance of the DVC is improved. To achieve this objective, this article deals with a hybrid SI generation scheme utilizing the principles of discrete wavelet transform (DWT) and extreme learning machine (ELM) algorithm in a transform domain-based DVC framework. Results: Exhaustive simulations have been carried out for some standard video sequences with the proposed and benchmark schemes. The proposed scheme is evaluated with respect to different performance metrics such as rate-distortion (RD), SI peak-signal-to-noise-ratio (PSNR) vs frame number, number of parity requests per SI frame, and so on. Experimental results and its analyses corroborate that the performance of the proposed technique surpasses as that of the benchmark schemes.