Text Image Watermark Detection Based on Improved YOLOv8 Model

The majority of contemporary watermark detection methodologies are founded upon the examination of natural images, with comparatively limited investigation into the domain of text images. The conventional detection techniques are not specifically designed to target text images, and they exhibit shortcomings such assuboptimal accuracy and low efficiency. To address the aforementioned issues, this paper conceptualises watermark detection as a target detection task and proposes a text image watermark detection method, YOLOv8-DCE, which enhances theYOLOv8 model. initially, the C2f module is augmented with the introduction of Dynamic Snake Convolution (DSConv), an innovation that builds upon the conventional convolution operation. The concept of a dynamic snake path is introduced into the convolutional operation. Secondly, a CA (Coordinate Attention) attention mechanism is incorporated to enhance the model's performance by learning the relationship between the channels. Finally, the bounding box loss function is modified to EloU (Efficient Intersection over Union), which improves the model's convergence speed and positioning accuracy. In order to assess the efficacy of the proposed watermark detection method on text images, a Text Image Watermark Dataset (TlWD) has been constructed. The experimental results demonstrate that the enhanced YOLOv8-DCE algorithm attains a Map value of 93.9% on the TlWD dataset, representing a 5.6 percentage point improvement over the traditional YOLOv8 algorithm and exhibiting superior detection capabilities.