Scholarlink

<p>In recent years, neural networks have typically relied on manual design, which relies on expert experience and are accompanied by high resource consumption. The emergence of Neural Architecture Search (NAS) has alleviated the need for expert experience in designing artificial neural networks. While existing NAS methods have achieved significant performance, the process of architecture search is often constrained by encoding schemes, search spaces, and optimization strategy, which lead to insufficient research in downstream applications. To address this issue, a joint optimization framework based on hybrid encoding and multi-objective Particle Swarm Optimization (PSO) is proposed. To obtain valuable initial architectures, a hybrid coding method of binary coding and Huffman coding is proposed to encode the parameters of the network modules respectively. A multi-objective optimization function is proposed to balance multiple conflicting parameters and search for the optimal architecture within parameter constraints. Additionally, we explore a valuable search space for automatically discovering high-performance neural network structures suitable for image classification and object detection. We conducted experiments on three classification datasets, CIFAR-10, CIFAR-100, and ImageNet. The proposed method was able to discover optimal architectures with only 0.3 GPU days of search cost, while achieving competitive performance on all benchmarks. We also applied the proposed search framework to the surface defect detection task of strip steel. The experimental results on NEU-DET and GC-DET datasets show that the proposed method achieves 81.4% and 74.5% on the mAP@0.5, outperforming several state-of-the-art detection approaches.</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

Hybrid encoding and multi-objective optimization-based neural architecture search for object detection

  • 期刊论文

  • 发表日期：

    2026年7月9日

查看全文
  • Liang Wang,
  • Yongjie Ma,
  • Tao Gong,
  • Quanxiu Li

摘要

In recent years, neural networks have typically relied on manual design, which relies on expert experience and are accompanied by high resource consumption. The emergence of Neural Architecture Search (NAS) has alleviated the need for expert experience in designing artificial neural networks. While existing NAS methods have achieved significant performance, the process of architecture search is often constrained by encoding schemes, search spaces, and optimization strategy, which lead to insufficient research in downstream applications. To address this issue, a joint optimization framework based on hybrid encoding and multi-objective Particle Swarm Optimization (PSO) is proposed. To obtain valuable initial architectures, a hybrid coding method of binary coding and Huffman coding is proposed to encode the parameters of the network modules respectively. A multi-objective optimization function is proposed to balance multiple conflicting parameters and search for the optimal architecture within parameter constraints. Additionally, we explore a valuable search space for automatically discovering high-performance neural network structures suitable for image classification and object detection. We conducted experiments on three classification datasets, CIFAR-10, CIFAR-100, and ImageNet. The proposed method was able to discover optimal architectures with only 0.3 GPU days of search cost, while achieving competitive performance on all benchmarks. We also applied the proposed search framework to the surface defect detection task of strip steel. The experimental results on NEU-DET and GC-DET datasets show that the proposed method achieves 81.4% and 74.5% on the mAP@0.5, outperforming several state-of-the-art detection approaches.