Scholarlink

Abstract <p>The article considers the problem of selecting the optimal forms and the number of anchor boxes for fine-tuning algorithms for detecting and classifying objects. The search for anchor boxes is reduced to a clustering problem that allows one to identify groups of similar objects. Three algorithms based on different principles were selected for clustering: the prototype principle, hierarchical trees, and graphs. For experiments with searching for anchor boxes, two neural network models were selected: SSD MobileNet V2 FPNLite 640 × 640 and SSD ResNet50 V1 FPN 640 × 640, found in the TensorFlow 2 Detection Model Zoo repository. These neural network models were pretrained on the Microsoft Common Objects in Context 2017 dataset and then further trained on the VisDrone2022 dataset. Using cluster analysis, 15 sets of aspect ratios were calculated. For each set of coefficients obtained, a separate neural network model SSD MobileNet V2 FPNLite 640 × 640 was trained and compared with the base models SSD MobileNet V2 FPNLite 640&#xa0;×&#xa0;640 and SSD ResNet50 V1 FPN 640 × 640 with standard values of the aspect ratios. As a result of optimization of parameters, the accuracy of object detection using a model based on the MobileNet V2 mobile neural network architecture has increased significantly. This result has great practical significance when working with compact energy-efficient systems with limited performance and memory.</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

Cluster Data Analysis for Optimization of Parameters of Object Detection Algorithms by Mobile Neural Network Models

  • 期刊论文

  • 发表日期：

    2025年6月25日

查看全文
  • P. O. Arkhipov,
  • S. L. Philippskikh,
  • M. V. Tsukanov

摘要

Abstract

The article considers the problem of selecting the optimal forms and the number of anchor boxes for fine-tuning algorithms for detecting and classifying objects. The search for anchor boxes is reduced to a clustering problem that allows one to identify groups of similar objects. Three algorithms based on different principles were selected for clustering: the prototype principle, hierarchical trees, and graphs. For experiments with searching for anchor boxes, two neural network models were selected: SSD MobileNet V2 FPNLite 640 × 640 and SSD ResNet50 V1 FPN 640 × 640, found in the TensorFlow 2 Detection Model Zoo repository. These neural network models were pretrained on the Microsoft Common Objects in Context 2017 dataset and then further trained on the VisDrone2022 dataset. Using cluster analysis, 15 sets of aspect ratios were calculated. For each set of coefficients obtained, a separate neural network model SSD MobileNet V2 FPNLite 640 × 640 was trained and compared with the base models SSD MobileNet V2 FPNLite 640 × 640 and SSD ResNet50 V1 FPN 640 × 640 with standard values of the aspect ratios. As a result of optimization of parameters, the accuracy of object detection using a model based on the MobileNet V2 mobile neural network architecture has increased significantly. This result has great practical significance when working with compact energy-efficient systems with limited performance and memory.