Scholarlink

With the advancement in visual communication, processes such as image analysis and orientation correction have become more crucial and critical than ever. The ability to correct image orientation is useful in many real-life applications like photography, forensic analysis, medical imaging, satellite imaging, robot-captured images, and many more. As per the literature, most of the works on image orientation correction have been achieved by bulky models which require heavy storage requirements. So there was a need to develop a lightweight model which could perform at par with the bulky models in accuracy and still consume less space making it deployable in low memory devices. This paper comes up with a architectural model capable of correcting image orientation in real time .The proposed model being light weight is capable of being deployed in mobile and embedded devices, also to harness the power of deep learning-based models this study uses a CNN model, pre-trained on ImageNet dataset and employed transfer learning for the feature extractor part coupled with fine-tuning techniques along with development of a novel classifier to aid the process of identifying the orientation of image among 270, 180, 90, and 0 degrees. In this study, the model was made to go through several datasets like MIT-Indoor, Inria Holidays, Coco dataset , Sun-397, and Pascal dataset and found that our model gave promising results.

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

Learning-Based Approach for Persistent Image Orientation Among Heterogeneous Domains

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Sandesh Verma,
  • Namita Tiwari,
  • Meenu Chawla,
  • Vishal Singh Bhati

摘要

With the advancement in visual communication, processes such as image analysis and orientation correction have become more crucial and critical than ever. The ability to correct image orientation is useful in many real-life applications like photography, forensic analysis, medical imaging, satellite imaging, robot-captured images, and many more. As per the literature, most of the works on image orientation correction have been achieved by bulky models which require heavy storage requirements. So there was a need to develop a lightweight model which could perform at par with the bulky models in accuracy and still consume less space making it deployable in low memory devices. This paper comes up with a architectural model capable of correcting image orientation in real time .The proposed model being light weight is capable of being deployed in mobile and embedded devices, also to harness the power of deep learning-based models this study uses a CNN model, pre-trained on ImageNet dataset and employed transfer learning for the feature extractor part coupled with fine-tuning techniques along with development of a novel classifier to aid the process of identifying the orientation of image among 270, 180, 90, and 0 degrees. In this study, the model was made to go through several datasets like MIT-Indoor, Inria Holidays, Coco dataset , Sun-397, and Pascal dataset and found that our model gave promising results.