Image Manipulation Localization Based on Multiscale Convolutional Attention

Current developments in image editing technology bring considerable challenges to the trustworthiness of multimedia data. Recent research has shown that convolutional attention mechanism can encode contextual information more effectively and efficiently than attentional mechanism, which has been validated in the field of semantic segmentation of images. However, convolutional attention-based networks focus only on the semantic information at the object level. The anomalous features introduced by tampering manipulation in the low-level information are ignored, resulting in poor localization of tampered regions. In this paper, leveraging convolutional attention mechanism, we propose an image tampering localization method based on multi-scale convolutional attention mechanism (IMLMCA), which fuses the low-level information to improve its capability to recognize the low-level anomalous information. To overcome the problems such as the unbalanced distribution of training positive and negative samples caused by the tampered regions, which usually occupy only a tiny part of the whole image, we introduce a mixed loss function consisting of focal-loss and Lovászloss to assist the method in learning the tampered features. The experimental results show that the proposed method achieves state-of-the-art performance on several publicly available tampered image datasets. For example, on CASIA dataset, the F1 score of the proposed method is 0.601, while the F1 score of the current best Transformer-based method, Objectformer, is 0.579. Codes are available at https://github.com/AlchemistLiu/IMLMCA .