【摘要】 针对真实场景下图像镜面高光降低图像质量的问题，提出了一个新颖的生成对抗网络(generative adversarial network, GAN)模型，可以有效去除单张真实场景图像中的高光。核心思想是通过VGG-19模块提取图像特征，然后利用GAN生成去除高光的图像，并对其质量进行判定。为了提高针对真实场景中不同尺度大小高光的去除能力，在生成器中加入了金字塔分割注意力机制和空间注意力机制，为多尺度的特征分配不同的权重，从而更聚焦于高光区域的特征，以提升纹理恢复效果。所训练的模型利用上下文损失函数、对抗损失函数和一致性损失函数来联合优化模型参数，提高了去除高光后图像与真实漫反射图像之间的相似性。在2个基准数据集与真实室内外场景中的测试实验结果表明，所提方法可以有效去除图像中的高光，并准确恢复高光区域的本征颜色与纹理细节。更多还原

【Abstract】 Aiming at the problem of reducing image quality by specular highlights in real-world scenes, a new generative adversarial network(GAN) model for specular highlights removal from real-world single images was proposed. To be specific, image features were extracted by using the VGG-19 module, then the proposed GAN was applied to generate specularity-free images and judge the quality of the generated images. In order to improve the removal ability for highlights with different scales in real-world scenes, a pyramid segmentation attention mechanism and a spatial attention mechanism were added to the generator. Different weights were assigned to multi-scale features, so as to focus more on features of highlight areas and improve the texture restoration effect. The trained model utilized the context loss function, adversarial loss function and consistency loss function to jointly optimize the model parameters, which improved the similarity between the image after highlight removal and the ground-truth diffuse image. Experimental results on two benchmark datasets and real-world images show that the proposed method can effectively remove the highlights and accurately restore the intrinsic color and texture details of the highlight areas.更多还原