A Hybrid Deep Learning Framework with Weighted Voting Ensemble of DenseNet and ResNet for watermark Text Classification

In recent years, the proliferation of digital con- tent has led to increased concerns about intellectual property protection and content authenticity. Digital watermarks, especially those embedded as text in scene images, have emerged as a vital tool for asserting ownership and preventing unauthorized use. However, detecting and classifying such watermarked text in complex natural scenes remains a challenging task due to varying backgrounds, fonts, distortions and noise. This research focuses on the classification of watermarked and non-watermarked text in scene im- ages. To achieve this, we used deep convolutional neural networks, DenseNet-169, Squeeze-and- Excitation (SE) -ResNet and Wide ResNet, for robust feature extraction. The extracted features were then passed to Machine Learning (ML) classifiers, including Random Forest (RF), Support Vector Ma- chine (SVM), and Logistic Regression (LR), to per- form the final classification. In addition, a weighted voting mechanism was implemented to combine the outputs of individual classifiers and improve over- all accuracy. The proposed approach demonstrates promising results in distinguishing between water- marked and non-watermarked text, contributing to the advancement of automated watermark detection systems. The proposed method shows promising results, with the highest accuracy of 97.13% achieved using SE- ResNet with weighted voting, followed by 95.08% using DenseNet201. This highlights SE-ResNet’s effectiveness in ensemble classification for watermark detection.