基于深度残差生成对抗网络的本征图像分解算法

AN INTRINSIC IMAGE DECOMPOSITION ALGORITHM BASED ON DEPTH RESIDUAL GENERATIVE ADVERSARIAL NETWORK

针对现有方法分解质量不佳、特征信息不够清晰的问题,提出一种基于深度残差生成对抗网络的本征图像分解算法,用于将单个图像本征分解为反照率和阴影分量。该算法是基于一个全卷积神经网络。通过引入残差块的单个端到端深序列以及两个经过对抗训练的判别器形成了对图像敏感的感知动机度量网络,在不需要任何物理先验和几何信息前提下,实现了单幅图像本征分解。实验结果表明,相对于其他算法,该方法具有更优的性能,而且获得对尺度敏感的反照率。

Aiming at the problem that the existing methods have poor decomposition quality and the feature information is not clear enough,an intrinsic image decomposition algorithm based on depth residual generative adversarial network is proposed to decompose intrinsically the single image into albedo and shadow component.The algorithm was based on a full-convolution neural network.By introducing a single end-to-end deep sequence of residual blocks and two confrontation-trained discriminators,an image-sensitive perceptual motivation measurement network was formed.Without any physical priori and geometric information,the intrinsic decomposition of a single image was realized.The experimental results show that the proposed method has better performance than other algorithms and can obtain scale-sensitive albedo.