Efficient Reconfigurable Mixed Precision ℓ₁ Solver for Compressive Depth Reconstruction
Published in Journal of Signal Processing Systems, 2022
This paper presents a mixed precision framework for ℓ₁ solvers, specifically focusing on the Alternating Direction Method of Multipliers (ADMM) and Proximal Gradient Descent (PGD) algorithms, applied to the Least Absolute Shrinkage and Selection Operator (LASSO) problem. The framework enables compressive depth reconstruction by varying the precision scaling in single-bit granularity during the iterative optimization process.
Key contributions include:
Mixed Precision Implementation: The ℓ₁ solvers are implemented with both floating-point and fixed-point arithmetic, allowing for fine-grained precision tuning during the iterative process.
FPGA Implementation: The solvers are deployed on an FPGA with a pipelined architecture, demonstrating the feasibility of the proposed approach in hardware.
Performance Evaluation: The implementation achieves over 74% savings in hardware resources and 60% in power consumption, with only minor reductions in reconstructed depth image quality compared to single-precision floating-point implementations.
The proposed approach provides a promising direction for efficient depth reconstruction in resource-constrained systems, such as embedded devices and mobile platforms.
Recommended citation: Wu, Y., Wallace, A. M., Mota, J. F. C., Aßmann, A., & Stewart, B. (2022). Efficient Reconfigurable Mixed Precision ℓ₁ Solver for Compressive Depth Reconstruction. *Journal of Signal Processing Systems*, 94(10), 1083–1099. https://doi.org/10.1007/s11265-022-01766-3
Download Paper | Download Slides | Download Bibtex