Article Preview
TopIntroduction
Image style transfer (IST) is designed to convert the textures, strokes, and colors of a content image to the textures, strokes, and colors of the style image. The scene structure and object boundary of the content image needs to be retained for a high-quality IST, and the appearance should be aligned with the style image. In recent years, IST technologies (Gatys et al., 2016; Ghiasi et al., 2017; Junginger et al., 2018; Karras et al., 2019; Li et al., 2018; Li & Wand, 2016; Liao & Huang, 2022; Qiao et al., 2021; Wang et al., 2020; Yao et al., 2019; Yeh et al., 2020; Zhang & Dana, 2017) based on deep learning (DL) have demonstrated that the relevance among features obtained by CNNs is very available for obtaining visual contents and styles, and is utilized to obtain images with similar contents and styles.
Generative adversarial networks (GANs) (Goodfellow et al., 2014) provide a new idea for image generation and a model basis for high-resolution image generation. GAN-based and convolutional neural networks (CNNs)-based methods have been broadly utilized in many fields, such as semantic web (Lv et al., 2020; Y. Zhu et al., 2017; Zhang et al., 2018), natural language processing (Alshdadi et al., 2021; Chui et al., 2022; Wang et al., 2017), image super resolution (Chen et al., 2021;Jiang et al., 2020), target detection (Qiu et al., 2021) and image segmentation (Jiang et al., 2019). At present, most current deep learning (DL)-based IST models have achieved relatively satisfactory results. However, they are insufficient to extract the texture features, leading to a lower definition of transferred style images. To address this issue, the authors focus on GAN-based methods and present a new IST method based on an enhanced GAN with a prior circular local binary pattern (LBP). The network structure of the generator is improved to enhance the capability of feature extraction and capture image semantic information in view of the instability of the generator output, and a more appropriate loss function is designed to enhance the visual effects. The main contributions are as follows:
- •
Circular LBP in a GAN generator is utilized as a texture prior to improving the detailed textures of the generated style image. Meanwhile, the authors integrate a dense connection residual block and an attention mechanism into the generator to further improve high-frequency feature extraction.
- •
The authors introduce the total variation (TV) regularizer into the loss function to smooth the training effects and restrain the noise while generating image contents.
- •
Experiments are conducted on Scenery2painting and MS-COCO datasets. The qualitative and quantitative results demonstrate that the metric quality of the generated images can achieve better effects by the authors’ proposed method compared with other outstanding approaches.
The rest of this paper is organized as follows. The authors present related work, followed by their proposed method. A performance evaluation is then presented, and finally a conclusion is given with a brief summary.