Scholarlink

<p>Embedded vision systems, particularly those deployed in autonomous vehicles, surveillance systems, and wearable devices, face significant challenges due to their stringent constraints on computational power, memory, energy consumption, and real-time processing requirements. To address these challenges, we proposed Adaptive Sparse Neural Architecture (ASNA), a lightweight and efficient computer vision algorithm specifically designed for resource-constrained environments. ASNA introduces four innovations: (1) Adaptive Sparsity with Dynamic Pruning, which optimizes computational efficiency by dynamically pruning neural network weights based on input complexity; (2) Hardware-Aware Quantization with Bit-Level Precision Control, enabling memory and energy savings through precision adjustments tailored to the target hardware; (3) Event-Driven Feature Extraction with Neuromorphic Computing, achieving ultra-low-latency processing by leveraging event-based data streams inspired by biological vision systems; and (4) Federated Learning for On-Device Model Personalization, ensuring scalability and privacy by enabling continuous model adaptation across distributed devices without sharing raw data. These innovations collectively enable ASNA to achieve state-of-the-art performance in tasks such as object detection, segmentation, and tracking while maintaining minimal resource usage. Experimental results on resource-constrained platforms such as NVIDIA Jetson Nano demonstrate that ASNA reduces computational overhead by up to 86.1%, memory usage by 50%, energy consumption by 58.4%, and latency by 90% compared to existing methods, making it ideal for applications in autonomous vehicles, surveillance systems, wearable devices, and industrial automation. By pushing the boundaries of efficiency and adaptability, ASNA represents a significant advancement in the field of embedded vision and sets a new benchmark for lightweight computer vision algorithms. The source code of the proposed ASNA is publicly available at github.com/livingjesus/ASNA.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

ASNA: A Novel Adaptive Sparse Neural Architecture for Embedded Systems

  • 期刊论文

  • 发表日期：

    2026年5月15日

查看全文
  • Idowu Paul Okuwobi,
  • Jingyuan Liu,
  • Olayinka Susan Raji,
  • Olusola Funsho Abiodun

摘要

Embedded vision systems, particularly those deployed in autonomous vehicles, surveillance systems, and wearable devices, face significant challenges due to their stringent constraints on computational power, memory, energy consumption, and real-time processing requirements. To address these challenges, we proposed Adaptive Sparse Neural Architecture (ASNA), a lightweight and efficient computer vision algorithm specifically designed for resource-constrained environments. ASNA introduces four innovations: (1) Adaptive Sparsity with Dynamic Pruning, which optimizes computational efficiency by dynamically pruning neural network weights based on input complexity; (2) Hardware-Aware Quantization with Bit-Level Precision Control, enabling memory and energy savings through precision adjustments tailored to the target hardware; (3) Event-Driven Feature Extraction with Neuromorphic Computing, achieving ultra-low-latency processing by leveraging event-based data streams inspired by biological vision systems; and (4) Federated Learning for On-Device Model Personalization, ensuring scalability and privacy by enabling continuous model adaptation across distributed devices without sharing raw data. These innovations collectively enable ASNA to achieve state-of-the-art performance in tasks such as object detection, segmentation, and tracking while maintaining minimal resource usage. Experimental results on resource-constrained platforms such as NVIDIA Jetson Nano demonstrate that ASNA reduces computational overhead by up to 86.1%, memory usage by 50%, energy consumption by 58.4%, and latency by 90% compared to existing methods, making it ideal for applications in autonomous vehicles, surveillance systems, wearable devices, and industrial automation. By pushing the boundaries of efficiency and adaptability, ASNA represents a significant advancement in the field of embedded vision and sets a new benchmark for lightweight computer vision algorithms. The source code of the proposed ASNA is publicly available at github.com/livingjesus/ASNA.