Hybrid Weighted Voting Framework for Watermark Text Classification Using Extended MobileNetV2 Model

The proliferation of digital media has made robust watermarking techniques essential for copyright protection and content authentication. There are many difficulties in detecting and classifying watermark text in images of nature scenes like variations in font, background complexity, and distortions such as perspective changes, occlusions, and lighting variations. This research proposes a hybrid approach for watermark text classification in natural scene images by integrating the Extended MobileNetV2 model with additional convolutional layers to enhance feature extraction with machine learning classifiers— Decision Tree (DT), Support Vector Machine (SVM), Random Forest (RF). The lightweight, due to its depthwise separable convolutions yet effective Extended MobileNetV2 model is used for feature extraction, while DT, SVM, and RF classifiers leverage these extracted features to learn the distinguishing characteristics of watermarked and non-watermarked text regions. This hybrid approach combines the feature extraction power of deep learning with the classification strengths of machine learning, achieving high accuracy in distinguishing watermarked text from non-watermarked text. The weighted ensemble of classifiers, combining predictions from an DT, SVM and RF through a weighted voting mechanism that prioritizes higher-performing classifiers, achieved an accuracy of 94.00%, a precision of 92.50%, a recall of 96.00%, and an F1-score of 94.20%. These results significantly outperform both the original Extended MobileNetV2 model’s built-in classification layer and Basic MobileNetV2, highlighting the benefits of our hybrid approach for robust watermark text classification.