In this work, we have explored the Xception architecture for classifying watermarked and non-watermark text in natural scene images. Leveraging depth-wise separable convolutions based on the Inception architecture, the Xception model achieves remarkable accuracy while utilizing fewer parameters. As a convolutional neural network, the Xception model excels in capturing intricate features from images, making it well-suited for discerning subtle differences between watermarked and non-watermarked text. During the training phase, the model requires a substantial dataset comprising both types of images to learn and recognize the distinguishing features effectively. Once trained, the Xception model exhibits high accuracy in classifying new images into their respective categories. Extensive experimentation of publically available standard datasets exhibits the effectiveness of the proposed approach. The performance is evaluated using a comprehensive range of metrics, including accuracy, precision, recall, and F1-score, resulting in respective values of 86.5%, 84.76%, 89%, and 86.82%.