SPNet: Seam carving detection via spatial-phase learning

Abstract

Seam carving is an image content-aware retargeting operation that can automatically insert seams to expand an image or remove seams to reduce image size. However, it can also perform illegal image tampering by inserting or removing objects. We observe that upsampling is a necessary step for seam removal or insertion, and cumulative them can lead to significant changes in the frequency domain, particularly in the phase spectrum. In fact, according to the properties of natural images, the phase spectrum retains rich frequency components, which can complement the loss of the amplitude spectrum and provide additional information. To this end, we propose a spatial phase-based network (SPNet) that combines spatial and phase spectra to capture retargeting artifacts for image seam carving detection. In addition, since the artifacts usually hide in the local regions for the seam carving operation, the local texture feature is more effective than the high-level semantic one. Based on this, we introduce a shallow network to reduce the receptive field, it can highlight the local features while suppressing high-level semantic information. Extensive experiments demonstrate that SPNet achieves state-of-the-art (SOTA) performance.