Scholarlink

<p>Detecting soiling on optical camera lenses using deep learning-based algorithms requires substantial datasets, which are challenging to acquire in large volumes at a low cost. Existing data generation techniques have limitations in fidelity, physical interpretability, and artistic input. We present a novel physics-based method for procedural generation of lens soiling data that seamlessly integrates with deep learning training pipelines. Our approach simulates the imaging process of lens soiling through physical optics and dissecting it into discrete stages. By harnessing fundamental image processing functions from OpenCV, we synthesize a variety of effects, including mud stains, water stains, water droplet refraction, and lens flare. This methodology ensures the effectiveness, interpretability, and high annotation accuracy of the generated data. The algorithms utilized are grounded in fundamental image processing functions, making them suitable for integration into standard deep learning training pipelines. Experiments demonstrate the efficacy of our generated data in the task of image classification on the real-world public dataset. In the experiments of soiling semantic segmentation, we also achieve improved results by adjusting the categories. Thereby, we validate its practical utility and potential for enhancing the performance of lens soiling detection systems. Relevant codes and data are published to <a href="https://github.com/JannLi/physical_lens_soiling">https://github.com/JannLi/physical_lens_soiling</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

Procedural generation of lens soiling data via physics-based simulation

  • 期刊论文

  • 发表日期：

    2025年6月21日

查看全文
  • Song Zhang,
  • Rui Zhang,
  • Jian Li,
  • Xuefeng Li

摘要

Detecting soiling on optical camera lenses using deep learning-based algorithms requires substantial datasets, which are challenging to acquire in large volumes at a low cost. Existing data generation techniques have limitations in fidelity, physical interpretability, and artistic input. We present a novel physics-based method for procedural generation of lens soiling data that seamlessly integrates with deep learning training pipelines. Our approach simulates the imaging process of lens soiling through physical optics and dissecting it into discrete stages. By harnessing fundamental image processing functions from OpenCV, we synthesize a variety of effects, including mud stains, water stains, water droplet refraction, and lens flare. This methodology ensures the effectiveness, interpretability, and high annotation accuracy of the generated data. The algorithms utilized are grounded in fundamental image processing functions, making them suitable for integration into standard deep learning training pipelines. Experiments demonstrate the efficacy of our generated data in the task of image classification on the real-world public dataset. In the experiments of soiling semantic segmentation, we also achieve improved results by adjusting the categories. Thereby, we validate its practical utility and potential for enhancing the performance of lens soiling detection systems. Relevant codes and data are published to https://github.com/JannLi/physical_lens_soiling.