Fully convolutional networks (FCNs) have been efficiently applied in splicing localization. However, the existing FCN-based methods still have three drawbacks: (a) their performance in detecting image details is unsatisfactory; (b) deep FCNs are difficult to train; (c) results of multiple FCNs are merged using fixed parameters to weigh their contributions. So, an improved method is proposed. Firstly, both the original spliced image and its corresponding residual image are regarded as the inputs of the network. Secondly, the residual block is introduced into FCN as residual-based FCN (RFCN) to make the network easier to optimize. Thirdly, three different RFCNs are merged to enhance locating maps with two learnable weight parameters. Besides, condition random field is introduced into the whole network to improve the results further. Experimental results on five datasets show that the proposed method performs better than some existing methods in localization ability, generalization ability, and robustness against additional operations.

全卷积网络（FCN）已有效地应用于拼接定位。但是，现有的基于FCN的方法仍然存在三个缺点：（a）它们在检测图像细节方面的性能不令人满意；（b）深层FCN很难训练；（c）使用固定参数合并多个FCN的结果以权衡其贡献。因此，提出了一种改进的方法。首先，原始拼接图像及其对应的残差图像均被视为网络的输入。其次，将残差块作为基于残差的FCN（RFCN）引入FCN，以使网络更易于优化。第三，将三个不同的RFCN合并以增强具有两个可学习的权重参数的定位图。此外，将条件随机字段引入整个网络以进一步改善结果。