Research on Lightweight and Fast Image Style Transform Algorithm

Image style transform technology refers to the use of a convolutional neural network to extract the style of a famous painting, thereby converting the input image into a corresponding style image. The current methods are mainly divided into two types, namely, the style transform method proposed by Gatys and the fast image style transform method (FNST) proposed by Li Feifei. The common problem between the two methods is that the hardware requirements are high, which is not conducive to a wide range of popularization and application. In order to use image style transform on more terminals and even mobile phones, research h on lightweight and fast image style transform is of great significance. This paper draws on the deep-wise split convolution in MobileNets to prun the convolutional layer in the loss network. After the pruning, the parameter content is only 11% of the original model; the residual component in Image Transform Net is trimmed. The parameter content was reduced by 96% compared to the original model. However, the amount of parameters in the process of weight reduction is greatly reduced, which may lead to the degradation of the stylized effect. It is proved by experiments that the style degradation of the method is not obvious, that is to say that the performance loss caused by the reduction of the parameter quantity is acceptable.