SPAN: Spatial Pyramid Attention Network forImage Manipulation Localization

TL;DR

SPAN introduces a multi-scale attention-based framework for detecting and localizing various image manipulations, leveraging a pyramid of local self-attention blocks with position encoding to improve accuracy.

Contribution

The paper proposes a novel Spatial Pyramid Attention Network (SPAN) that models multi-scale patch relationships for image manipulation localization, outperforming previous methods.

Findings

Significant performance improvements over state-of-the-art methods.

Effective modeling of multi-scale patch relationships.

Flexible training on synthetic data with fine-tuning capability.

Abstract

We present a novel framework, Spatial Pyramid Attention Network (SPAN) for detection and localization of multiple types of image manipulations. The proposed architecture efficiently and effectively models the relationship between image patches at multiple scales by constructing a pyramid of local self-attention blocks. The design includes a novel position projection to encode the spatial positions of the patches. SPAN is trained on a generic, synthetic dataset but can also be fine tuned for specific datasets; The proposed method shows significant gains in performance on standard datasets over previous state-of-the-art methods.