Comparative Evolution of Digital Image Forgery Detection: From Residual-Saliency Processing to Variational Attention Learning

Abstract

Digital image forgery detection has progressively shifted from artifact preservation and suspicious-region segmentation toward richer feature learning and uncertainty-aware decision making. This article presents a comparative analysis of two sequential research frameworks developed for image forgery analysis. The first framework combines Dynamic Residual Forgery-aware Noise Suppression (DRFNS) with Gradient-Aware Patch Saliency (GAPS) to suppress noise while retaining forensic residuals and to segment suspicious regions. The second framework extends the analysis through Forgery-Aware Semantic and Edge Tracing Feature Extraction (FASETE) and a Variational Evidence-Driven Attention Classifier (VEDAC), integrating multi-domain feature representation, attention refinement, variational inference, confidence estimation and patch-level localization. Using the described experimental results from the two studies, the comparison examines the evolution of preprocessing, feature representation, localization and classification capabilities. The analysis shows that the research evolution moves from evidence-preserving localization in Phase I toward integrated feature learning and uncertainty-aware classification in Phase II. The study also identifies the remaining need for cross-dataset validation, stronger recall and evaluation under unseen post-processing conditions.