Image Manipulation Localization Using Spatial–Channel Fusion Excitation and Fine-Grained Feature Enhancement

The purpose of image manipulation detection is to classify and locate tampered regions in digital images. Most existing manipulation localization methods usually rely on certain tampering traces hidden in manipulated images. This dependency, however, may damage the generalization and postprocessing capabilities of the detection model because the tampered content in the image may be weakened by postprocessing operations. To address the aforementioned problem, we propose a new image manipulation localization scheme by introducing spatial–channel fusion excitation and fine-grained feature enhancement (FFE). We first design a feature enhancement module to enhance fine-grained features in red green blue (RGB) streams, which can improve the localization accuracy of tampering regions by capturing different-scale local and global information of images. Furthermore, a fusion excitation strategy is introduced to efficiently fuse features from both spatial and channel domains. Our fusion strategy can simultaneously process image spatial and channel information, significantly enhancing the model's differentiation capability between tampered and nontampered regions. Extensive experiments demonstrate that the proposed method can provide effective localization capability for multiscale manipulation regions over different image sets and outperform most of the state-of-the-art schemes in terms of detection accuracy, generalization, and robustness.