Visible Watermark Detection Based on an Improved YOLOv8 Model

Abstract

With the widespread dissemination of digital images on the internet, issues related to image copyright protection and content security have become increasingly prominent, making visible watermarks a common method of copyright marking. To address the frequent inaccuracies in visible watermark detection under complex background conditions, this paper proposes an improved deep learning-based visible watermark detection model that achieves high-precision detection and localization of watermark regions. structural optimization is conducted based on the YOLOv8 object detection framework. Considering the characteristics of visible watermarks, such as large scale variations, uneven transparency, and strong coupling with background textures, an attention mechanism module is introduced into the backbone network to enhance feature representation of critical regions. In addition, a watermark feature enhancement module is designed to strengthen multi-scale feature fusion. Furthermore, the improved YOLOv8 model is fused with the RT-DETR model, combining the local feature extraction capability of convolutional neural networks with the global modeling ability of the Transformer architecture to improve detection accuracy and localization robustness. Experimental results demonstrate that the proposed method achieves an mAP50 of 98.9% and an mAP50:95 of 97.6% on the LVW dataset, outperforming multiple mainstream detection models and effectively balancing detection accuracy with efficiency.