A Gaussian Mixture Model Approach for Robust Watermark Text Detection in Documents

Abstract

Watermark text detection is a crucial process in image processing and digital forensics, particularly for ensuring content authenticity and preventing unauthorized use of digital media. Watermarks, often embedded in images, serve as a visible or invisible form of protection, and their detection is essential for verifying the integrity and ownership of digital assets. Detecting watermark text which is often subtle, semi-transparent, or integrated into the background presents significant challenges. In this paper, we propose an effective approach for watermark text detection using a Gaussian Mixture Model (GMM), an edge detection technique, and morphological functions. A Gaussian Mixture Model (GMM) with four cluster components is applied to model the distribution of pixel intensities. The GMM represents the intensity histogram as a mixture of Gaussian distributions, each parameterized by its mean, variance, and weight. Canny edge detection is applied to the grayscale image to retain the actual edge information. The resulting image is then refined using morphological closing a process that involves dilation followed by erosion to fill small gaps and smooth edge contours. To further isolate meaningful edges, Otsu’s thresholding method is applied, selecting an optimal threshold that minimizes intra-class variance between the foreground and background. Finally, the watermark text is embedded into the image by modifying the pixel values in a predefined region, resulting in a watermarked output image.