An image manipulation localization method based on multi-scale frequency awareness

With the rapid development of image editing techniques, accurately localizing manipulated regions remains challenging due to subtle artifacts and diverse manipulation patterns. Existing methods insufficiently exploit multi-scale frequency cues and often lack effective cross-level feature interaction. To address this, we propose a multi-scale frequency-aware framework for reliable manipulation localization. The core MFAFF module introduces adaptive low-pass and high-pass filtering to suppress irrelevant noise in high-level features while enhancing fine boundary details in low-level features. A lightweight depthwise separable convolutional decoder further improves spatial modeling efficiency. A joint loss combining BCE, Dice, and edge supervision enhances region accuracy and boundary precision. Extensive cross-dataset experiments demonstrate the effectiveness of adaptive frequency modeling in robust manipulation localization.