Scholarlink

<p>In the present era of multimedia communication, image security continues to be a challenging problem. In recent times, the image-hashing approach has drawn a lot of attention due to its ability to secure multimedia content by generating hash. This paper introduces an innovative image hashing approach that utilizes Siamese and Inception-ResNet-v2 networks to generate the image hash for checking the similarity between original and tampered images. The proposed deep hashing network learns to automatically extract features in alignment with training objectives, ending in the generation of the final hash. By incorporating stem, Inception-ResNet, reduction layers, and global average pooling layers, the size of the input image is reduced while maintaining enhanced channel depth. This technique dynamically adjusts the organization of the training set in response to variations in constraint values. Empirical outcomes suggest that the proposed deep hashing network achieves better performance in terms of robustness and discrimination. Comprehensive evaluations and receiver operating characteristic curve measured on a large test dataset demonstrate our method’s superiority in terms of content authentication and tampered detection over different state-of-the-art approaches.</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

Inception-ResNet-v2 Siamese Network-Based Image Hashing Technique for Tampering Detection

  • 期刊论文

  • 发表日期：

    2025年12月3日

查看全文
  • Abdul Subhani Shaik,
  • Ram Kumar Karsh,
  • Mohiul Islam,
  • D. N. Kiran Pandiri

摘要

In the present era of multimedia communication, image security continues to be a challenging problem. In recent times, the image-hashing approach has drawn a lot of attention due to its ability to secure multimedia content by generating hash. This paper introduces an innovative image hashing approach that utilizes Siamese and Inception-ResNet-v2 networks to generate the image hash for checking the similarity between original and tampered images. The proposed deep hashing network learns to automatically extract features in alignment with training objectives, ending in the generation of the final hash. By incorporating stem, Inception-ResNet, reduction layers, and global average pooling layers, the size of the input image is reduced while maintaining enhanced channel depth. This technique dynamically adjusts the organization of the training set in response to variations in constraint values. Empirical outcomes suggest that the proposed deep hashing network achieves better performance in terms of robustness and discrimination. Comprehensive evaluations and receiver operating characteristic curve measured on a large test dataset demonstrate our method’s superiority in terms of content authentication and tampered detection over different state-of-the-art approaches.