Efficient Approximate Checkerboard K-SVD for Resource Constrained Embedded Systems
Date:
This talk presented a checkerboard-style approximate K-SVD algorithm tailored for resource-constrained embedded systems.
Key contributions included:
- Analysis of resource-constrained environments (size, weight, power, cost)
- Development of checkerboard K-SVD using approximate linear algebra (SXLAL library)
- Evaluation of compressed image quality via PSNR and SSIM
- Comparison of floating-point, fixed-point, and posit (Unum) arithmetic
- Demonstrated up to 16× reduction in computational complexity while maintaining image quality
- Achieved up to 98% memory footprint reduction using reduced precision
The work showcased how approximate computing strategies can enable efficient sparse representation in embedded vision applications.
