This paper addresses the challenge of classifying watermarked and non-watermark text in natural scene images by leveraging the MobileNetV2 architecture. MobileNetV2, known for its lightweight design and efficiency, is extended and customized to enhance performance for this specific task. The proposed model incorporates additional layers and employs transfer learning to effectively adapt pre-trained knowledge, achieving high accuracy even with limited data. Key enhancements, such as a global average pooling layer and a sigmoid activation function, enable robust feature extraction and binary classification. Experimental results validate the model’s effectiveness, achieving 92.44% accuracy, 89.96% precision, 95.56% recall, and 92.67% F1-score, highlighting its potential for real-time applications in watermark text classification.