PL-GNet: Pixel Level Global Network for detection and localization of image forgeries

Zenan Shi, Xuanjing Shen, Haipeng Chen, Yingda Lyu

Abstract: Highlights•A pixel-level global network that considers different kinds of inputs is proposed to localize manipulated regions in the image.•A new first layer and the backbone architecture in Encoding net are designed to extract efficient and discriminative features.•In addition, the sub-network LSTM Network, which introduces co-occurrence based features as input, is designed in the frequency domain.•Our PL-GNet is not limited to one specific manipulation type and shows superior performance on six challenging datasets.