The forgery of watermark images poses a significant threat to digital works’ visual integrity and authenticity. This research focuses on developing a method for detecting forged watermark images by comparing the performance of two conventional approaches: the Gabor and the Prewitt filters. In addition, a Convolutional Neural Network (CNN) is employed as a classifier to model complex features that may be challenging to identify using traditional detection methods. The Gabor and Prewitt filters extract texture and edge features from the images. Experiments were conducted on a dataset containing various forged watermark images with different levels of complexity. Performance evaluation results were measured by assessing accuracy during validation and testing, True Positive Rate (TPR), False Negative Rate (FNR), and the ROC–AUC curve. The research findings indicate that the Prewitt filter outperforms the Gabor filter, achieving a testing accuracy of 88.32%, a TPR of 82.94%, an FNR of 17.05%, and an ROC–AUC value of 0.884 for validation accuracy. These results demonstrate the potential of the Prewitt filter and CNN as a practical approach to addressing the challenges of detecting increasingly complex watermark image forgeries. This study contributes to developing watermark image authenticity detection methods by combining conventional approaches with deep learning technology.