Scholarlink

Abstract <p>The algorithm for searching for small-sized objects against the background of a non-uniform random texture proposed earlier by the author has been modified. The modification allows working in parallel with several segmentations of the image. The weights searched by training based on the error backward propagation (backpropagation) are attributed to the segmentations. In the realized process, five weights have been determined from the set consisting of a thousand photographic images of clouds with a size of <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(1200\times 1200\)</EquationSource> <!--OptelIns2570049Likhachev-m1--> </InlineEquation> pixels onto which model objects—circles with a radius of 2 pixels—were marked. Initially, all the weights were identical and equal to 0.2. After training, the maximum and minimum of them appeared to be 0.404 and 0.116, respectively. The graphs showing the changes in the weights are evidence that the training procedure converges. According to the results of the performed computation experiment, the modification has proved to be more efficient than the initial algorithm: the number of errors of the first and second kind decreased by 1.23 and 1.8 times, respectively</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

Training Algorithm of Searching for Small-Sized Objects in an Image

  • 期刊论文

  • 发表日期：

    2025年8月1日

查看全文
  • A. V. Likhachev

摘要

Abstract

The algorithm for searching for small-sized objects against the background of a non-uniform random texture proposed earlier by the author has been modified. The modification allows working in parallel with several segmentations of the image. The weights searched by training based on the error backward propagation (backpropagation) are attributed to the segmentations. In the realized process, five weights have been determined from the set consisting of a thousand photographic images of clouds with a size of \(1200\times 1200\) pixels onto which model objects—circles with a radius of 2 pixels—were marked. Initially, all the weights were identical and equal to 0.2. After training, the maximum and minimum of them appeared to be 0.404 and 0.116, respectively. The graphs showing the changes in the weights are evidence that the training procedure converges. According to the results of the performed computation experiment, the modification has proved to be more efficient than the initial algorithm: the number of errors of the first and second kind decreased by 1.23 and 1.8 times, respectively