Scholarlink

<p>A logo symbolises the identity of a brand or an organisation. It instils trust and creates a loyal community of customers. However, this trust is often exploited by counterfeiters who forge these logos for their malicious intent. Although the field of computer vision has made notable advances, a dearth of standardised benchmark datasets for counterfeit logo detection impedes progress in this discipline. Therefore, this study aims to address this gap by creating a synthetic counterfeit logo dataset using Stable Diffusion. Furthermore, it is paramount to develop automated mechanisms for detecting this type of forgery. Therefore, this study aims to analyse the performance of the ResNet variants, both with and without transfer learning, on counterfeit logo detection. The results demonstrate that models harnessing transfer learning outperform their non-pretrained counterparts, with ResNet-101 showing the most promising result of 0.97 AUC.</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

Leveraging Transfer Learning in ResNet Variants for Logo Infringement Detection

  • 期刊论文

  • 发表日期：

    2025年12月29日

查看全文
  • Pratishtha Gupta,
  • Bhawna Narwal,
  • Amar Kumar Mohapatra

摘要

A logo symbolises the identity of a brand or an organisation. It instils trust and creates a loyal community of customers. However, this trust is often exploited by counterfeiters who forge these logos for their malicious intent. Although the field of computer vision has made notable advances, a dearth of standardised benchmark datasets for counterfeit logo detection impedes progress in this discipline. Therefore, this study aims to address this gap by creating a synthetic counterfeit logo dataset using Stable Diffusion. Furthermore, it is paramount to develop automated mechanisms for detecting this type of forgery. Therefore, this study aims to analyse the performance of the ResNet variants, both with and without transfer learning, on counterfeit logo detection. The results demonstrate that models harnessing transfer learning outperform their non-pretrained counterparts, with ResNet-101 showing the most promising result of 0.97 AUC.