Wavelet-driven Partial Binary Quantization for accelerating deep-learning-based spectrum sensing in embedded systems

Qiao Sun

doi:10.1016/j.sysarc.2026.103842

Wavelet-driven Partial Binary Quantization for accelerating deep-learning-based spectrum sensing in embedded systems

2026-05-30

0

PRE

AI

DOI：10.1016/j.sysarc.2026.103842

Original

Original request for help

Save

Abstract

En 中文

Emerging cognitive radio applications in IoT and other resource-constrained embedded systems demand compact yet accurate deep-learning-based spectrum sensing models. However, conventional model compression methods often suffer from severe performance degradation because they overlook the unique frequency-domain characteristics of spectrum data. In this work, we propose a Wavelet-based Partial Binary Quantization (WPBQ) method, which evaluates kernel importance by jointly exploiting wavelet energy and cosine similarity so that frequency-critical information can be better preserved during partial binarization. To support efficient deployment, we further design SSAcc, an FPGA accelerator with heterogeneous GEMMFixed and GEMMBin engines to collaboratively execute fixed-point and binary computations. Experimental results show that WPBQ achieves an 8.2× compression ratio while maintaining near-full-precision accuracy, and it consistently outperforms representative binary quantization methods such as XNOR-Net and Bi-Real on the spectrum sensing task. Compared with existing binary quantization methods, WPBQ improves accuracy by up to 4.6%–19.9% under different binarization settings. The proposed accelerator further provides substantial speedup and memory reduction over typical embedded processors, demonstrating the effectiveness of the proposed hardware–software co-design for real-time embedded spectrum sensing.

Wavelet-driven Partial Binary Quantization for accelerating deep-learning-based spectrum sensing in embedded systems

Abstract

Journal

Organization