Open Access

ARTICLE

An Investigation of Frequency-Domain Pruning Algorithms for Accelerating Human Activity Recognition Tasks Based on Sensor Data

by Jian Su¹, Haijian Shao^1,2,*, Xing Deng¹, Yingtao Jiang²

1 School of Computer, Jiangsu University of Science and Technology, Zhenjiang, 212003, China
2 Department of Electrical and Computer Engineering, University of Nevada, Las Vegas, NV 89154, USA

* Corresponding Author: Haijian Shao. Email:

Computers, Materials & Continua 2024, 81(2), 2219-2242. https://doi.org/10.32604/cmc.2024.057604

Received 22 August 2024; Accepted 15 October 2024; Issue published 18 November 2024

Abstract

The rapidly advancing Convolutional Neural Networks (CNNs) have brought about a paradigm shift in various computer vision tasks, while also garnering increasing interest and application in sensor-based Human Activity Recognition (HAR) efforts. However, the significant computational demands and memory requirements hinder the practical deployment of deep networks in resource-constrained systems. This paper introduces a novel network pruning method based on the energy spectral density of data in the frequency domain, which reduces the model’s depth and accelerates activity inference. Unlike traditional pruning methods that focus on the spatial domain and the importance of filters, this method converts sensor data, such as HAR data, to the frequency domain for analysis. It emphasizes the low-frequency components by calculating their energy spectral density values. Subsequently, filters that meet the predefined thresholds are retained, and redundant filters are removed, leading to a significant reduction in model size without compromising performance or incurring additional computational costs. Notably, the proposed algorithm’s effectiveness is empirically validated on a standard five-layer CNNs backbone architecture. The computational feasibility and data sensitivity of the proposed scheme are thoroughly examined. Impressively, the classification accuracy on three benchmark HAR datasets UCI-HAR, WISDM, and PAMAP2 reaches 96.20%, 98.40%, and 92.38%, respectively. Concurrently, our strategy achieves a reduction in Floating Point Operations (FLOPs) by 90.73%, 93.70%, and 90.74%, respectively, along with a corresponding decrease in memory consumption by 90.53%, 93.43%, and 90.05%.

---

Keywords

---