基于PoolNet的改良风格迁移算法

A Modified Style Transfer Algorithm Based on PoolNet

随着深度学习的深入研究,将一张照片转变成具有明显的艺术家风格的画作的技术已经初步实现。然而,现有的大多数任务所生成的风格化图像效果并不佳,例如风格化图像的显著性区域较源图像发生了很大的变化、风格化图像中显著性强的区域出现光圈和伪影的情况。因此论文提出一种基于图像显著性的改良风格迁移算法,引入内容图像与风格图像之间的显著性损失来解决这一问题。此外,由于在计算特征相关性时,往往要求特征图拥有较大的感受野,因此普遍的做法是增加神经网络的层数,这会导致网络的参数量成倍增加,影响训练的速率。论文通过将网络结构中传统残差模块替换为转置残差模块,大大减少了网络的参数量,降低了对计算资源的要求。通过实验验证,与其他图像风格迁移算法相比,该算法不仅可以生成风格化效果更佳的画作图像,而且大大减少了训练时间。

With intensive research in deep learning,techniques for transforming a photograph into a painting with a distinct artist's style have been initially implemented.However,most of the existing tasks generate stylised images with poor results,for example,the saliency region of the stylised image changes significantly compared to the source image,and the aperture and artefacts appear in the salient region of the stylised image.Therefore,this paper proposes a modified style transfer algorithm based on image saliency,introducing saliency loss between the content image and the stylised image to solve this problem.In addition,since feature maps are often required to have large perceptual fields when calculating feature correlation,it is common practice to increase the number of layers of a neural network,which can lead to an exponential increase in the number of parameters in the network and affect the rate of training.In this paper,by replacing the traditional residual modules with Inverted residual modules in the network structure,the number of parameters of the network is greatly reduced and the requirement for computational resources is reduced.Through experimental validation,the algorithm not only generates better stylised images of paintings compared to other image style transfer algorithms,but also greatly reduces the training time.