Scholarlink

In this paper the improved algorithm for detecting and classifying objects based on the of a transformer architecture with the Flash Attention mechanism and a dataset of archaeological sites of the Bronze Age in the Southern Trans-Urals was proposes. Dataset contains remote sensing data and high-resolution aerial photo-graphs. The proposed algorithm allows detecting the location of ancient burial mounds and settlements with high accuracy, while eliminating the main drawback of the Flash Attention mechanism, associated with the high computational complexity. Low-altitude aerial photography generates a huge data array, so archaeologists face an important task of selecting frames containing signs of an archaeological sites. In this case, not only quality metrics are important, but also an algorithmic complexity of the classification procedure is of great importance. The results obtained based on various YOLO versions (v8, v11, v12) are presented and discussed, and the advantages of the proposed algorithm based on the Block-Sparse Flash Attention are shown.

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

Remote Sensing Detection an Archaeological Sites of Bronze Age Using Digital Terrain Models and YOLO Models Based on a Transformer Architecture

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • A. V. Vokhmintcev,
  • V. R. Abbazov,
  • M. M. A. Khater,
  • M. A. Romanov,
  • T. S. Vokhmintceva

摘要

In this paper the improved algorithm for detecting and classifying objects based on the of a transformer architecture with the Flash Attention mechanism and a dataset of archaeological sites of the Bronze Age in the Southern Trans-Urals was proposes. Dataset contains remote sensing data and high-resolution aerial photo-graphs. The proposed algorithm allows detecting the location of ancient burial mounds and settlements with high accuracy, while eliminating the main drawback of the Flash Attention mechanism, associated with the high computational complexity. Low-altitude aerial photography generates a huge data array, so archaeologists face an important task of selecting frames containing signs of an archaeological sites. In this case, not only quality metrics are important, but also an algorithmic complexity of the classification procedure is of great importance. The results obtained based on various YOLO versions (v8, v11, v12) are presented and discussed, and the advantages of the proposed algorithm based on the Block-Sparse Flash Attention are shown.