【摘要】 为解决现有局部生成式对抗网络(GAN)生成人脸检测算法在检测经过后处理的图像时性能严重下降的问题，提出一种注意力融合双流特征的局部GAN生成人脸检测算法.该算法利用双流网络分别从RGB颜色空间和YCbCr颜色空间中提取鲁棒特征，并引入注意力特征融合模块在不同网络层上融合双流特征以获得更鲁棒的特征.同时采用多层次特征融合决策提高网络对局部生成区域特征的提取和辨别能力.实验结果表明，所提算法的鲁棒性优于现有算法，尤其是针对JPEG压缩和双边滤波后处理.在FFHQ+规则子集上与次优算法相比，该算法在3种强度的JPEG压缩和双边滤波上的平均准确率分别提高了1.88%和2.64%;在FFHQ+不规则子集上与次优算法相比，该算法在3种强度的JPEG压缩和双边滤波上的平均准确率分别提高了2.85%和1.60%.更多还原

【Abstract】 In order to solve the problem that the performance of existing locally generative adversarial network(GAN) generated face detection algorithms degrades seriously when detecting post-processed images, a locally GAN-generated face detection algorithm based on dual-stream features fused by attention is proposed. The algorithm uses a dual-stream network to extract robust features from RGB color space and YCbCr color space, and an attentional feature fusion module is introduced to fuse dual-stream features on different network layers to obtain more robust features. Then, the decision is made by fusing multi-level features to improve the feature extraction and distinction ability of the network from locally generated region. Experimental results show that the robustness of the proposed algorithm is better than that of existing algorithms, especially for joint photographic experts group(JPEG) compression and bilateral filtering post-processing. Compared with the suboptimal algorithm tested on the Flickr faces high quality(FFHQ)+regular subset, the average accuracy of the algorithm for three intensities of JPEG compression and bilateral filtering is improved by 1.88% and 2.64%, respectively. Compared with the suboptimal algorithm tested on the FFHQ+irregular subset, the average accuracy of the algorithm for three intensities of JPEG compression and bilateral filtering is improved by 2.85% and 1.60%, respectively.更多还原