Prediction-Based Incremental Refinement for Binomially-Factorized Discrete Wavelet Transforms

摘要: It was proposed recently that quantized representations of the input source (e.g., images, video) can be used for computation two-dimensional discrete wavelet transform (2D DWT) incrementally. The coarsely is initial forward or inverse DWT, and result successively refined with each new refinement description via an embedded quantizer. This based on direct factorization DWT using generalized spatial combinative lifting algorithm. In this correspondence, we investigate use prediction results, i.e., exploiting correlation neighboring samples (or coefficients) in order to reduce dynamic range required computations, thereby circuit activity arithmetic operations transform. We focus binomial factorizations DWTs include (amongst others) popular 9/7 filter pair. Based FPGA co-processor testbed, present energy-consumption results incremental prediction-based comparison conventional (nonrefinable) computation. Our tests combinations intra error frames video sequences show former 70% more energy efficient than latter computing half precision remains 15% full-precision