Error Correcting Schemes for a Low-power Approximate DCT Design

Applications such as audio processing, image processing, and data mining do not always require strictly precise calculations. Recently, approximate computing has gained popularity as a promising technique for improving circuit performance in terms of delay, area, and power. JPEG is one example that benefits from this property. We are studying the application of approximate computing techniques to JPEG circuits in order to improve their power efficiency. Unfortunately, however, our previous results revealed that applying approximate computing techniques to JPEG sometimes has limited effectiveness and can lead to increased power consumption, contrary to expectations. This paper proposes error correcting schemes to address this problem, based on an analysis of its causes. Experimental evaluations show up to a 30% reduction in power consumption while maintaining image quality.