Scholarlink

<p>Digital images are widely used in modern life as one of the most prominent forms of communication and information medium. The cost and level of competence needed to take advantage of image tampering have significantly decreased due to the widespread availability of different photo editing software. Splicing is a form of forgery on digital images where segments of two or more pictures are combined to create a new synthetic image, many times aimed towards spreading of false propaganda through social media platforms, hence often misleading individuals or certain communities. This work proposes a novel deep learning-based technique utilizing the Multiscale Attention Network and the ResNext architecture to detect spliced region(s) in a forged image efficiently. Our enhanced model leverages the multiscale information and attention mechanism and ResNext’s powerful feature extraction capabilities. With this integration, the network can handle complex forging tasks like varying picture sizes and quality and improve splicing localization accuracy. Through comprehensive experiments on benchmark datasets, we have shown that our model consistently achieves an impressive accuracy of over <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(90\%\)</EquationSource> </InlineEquation> and an AUC score <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(&gt; 0.95\)</EquationSource> </InlineEquation>, surpassing current state-of-the-art methods in localizing spliced regions in images. The performance analysis of our proposed model against different post-processing techniques, including JPEG compression, proves our model’s superior robustness.</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 multiscale attention network model utilizing ResNext architecture for detection and localization of image splicing attack

  • 期刊论文

  • 发表日期：

    2025年7月25日

查看全文
  • Debjit Das,
  • Debolina Ghosh,
  • Aryan Raj,
  • Rupayan Thakur Chakraborty,
  • Anoushka Patra,
  • Ruchira Naskar

摘要

Digital images are widely used in modern life as one of the most prominent forms of communication and information medium. The cost and level of competence needed to take advantage of image tampering have significantly decreased due to the widespread availability of different photo editing software. Splicing is a form of forgery on digital images where segments of two or more pictures are combined to create a new synthetic image, many times aimed towards spreading of false propaganda through social media platforms, hence often misleading individuals or certain communities. This work proposes a novel deep learning-based technique utilizing the Multiscale Attention Network and the ResNext architecture to detect spliced region(s) in a forged image efficiently. Our enhanced model leverages the multiscale information and attention mechanism and ResNext’s powerful feature extraction capabilities. With this integration, the network can handle complex forging tasks like varying picture sizes and quality and improve splicing localization accuracy. Through comprehensive experiments on benchmark datasets, we have shown that our model consistently achieves an impressive accuracy of over \(90\%\) and an AUC score \(> 0.95\) , surpassing current state-of-the-art methods in localizing spliced regions in images. The performance analysis of our proposed model against different post-processing techniques, including JPEG compression, proves our model’s superior robustness.