Scholarlink

<p>Recognizing thin outdoor structures under complex illumination is challenging, especially when collecting large annotated datasets is impractical. This motivates the one-shot setting, aiming to recognize a class of objects from just a single annotated exemplar. We address this scenario for vision-based multimedia systems deployed outdoors with a unified, real-time framework. We propose LITE, an ultra-lightweight enhancement network that estimates iterative brightness adjustment curves to restore contrast and fine details; it runs at 1000&#xa0;fps on 2K images with 0.19K parameters. We further develop MSIO-SAM, a one-shot segmentation method that adapts the Segment Anything Model using multi-scale similarity analysis, clustering-based keypoint prompting, and iterative mask refinement to localize thin, low-contrast targets from a single annotated reference. Across diverse outdoor scenes, the pipeline improves IoU over SAM variants and reduces false negatives while remaining edge-deployable thanks to its small footprint and high throughput. The design applies to power-line inspection, aerial surveillance, and intelligent transportation, where robust one-shot recognition under challenging illumination is critical. Code is released at <a href="https://github.com/mubaisam/One-Shot-Power-Line-Localization.">https://github.com/mubaisam/One-Shot-Power-Line-Localization.</a></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

A unified lightweight enhancement–segmentation framework for one-shot recognition in complex outdoor lighting

  • 期刊论文

  • 发表日期：

    2026年5月6日

查看全文
  • Guangrui Bai,
  • Yahui Deng,
  • Diya He,
  • Yifan Mei,
  • Yanghong Li,
  • Yuhan Chen,
  • Erbao Dong

摘要

Recognizing thin outdoor structures under complex illumination is challenging, especially when collecting large annotated datasets is impractical. This motivates the one-shot setting, aiming to recognize a class of objects from just a single annotated exemplar. We address this scenario for vision-based multimedia systems deployed outdoors with a unified, real-time framework. We propose LITE, an ultra-lightweight enhancement network that estimates iterative brightness adjustment curves to restore contrast and fine details; it runs at 1000 fps on 2K images with 0.19K parameters. We further develop MSIO-SAM, a one-shot segmentation method that adapts the Segment Anything Model using multi-scale similarity analysis, clustering-based keypoint prompting, and iterative mask refinement to localize thin, low-contrast targets from a single annotated reference. Across diverse outdoor scenes, the pipeline improves IoU over SAM variants and reduces false negatives while remaining edge-deployable thanks to its small footprint and high throughput. The design applies to power-line inspection, aerial surveillance, and intelligent transportation, where robust one-shot recognition under challenging illumination is critical. Code is released at https://github.com/mubaisam/One-Shot-Power-Line-Localization.