Facial sketch synthesis (FSS) has advanced significantly in recent years, but challenges remain in few-shot settings. Some few-shot learning methods can convert photos (source domain) into sketches of a specified style (target sketch domain). However, they overlook the available samples of other sketch styles (non-target sketch domains). We argue that the information in these samples can help the model enhance its mapping ability from the source domain to the target domain. This paper proposes a progressive domain gap reduction (PDGR) method for few-shot facial sketch synthesis, which consists of three stages: teacher training, knowledge distillation, and intra-domain few-shot adaptation. In the first stage, we adapt a pretrained StyleGAN to a non-target sketch domain with more available samples than the target sketch domain. To generate diverse and high-quality sketches, we employ a dual-discriminator adversarial mechanism to guide the model in focusing on the overall structure and style, as well as multi-scale details and textures. In the second stage, the knowledge from StyleGAN is transferred to a U-Net for more efficient image translation. In the third stage, we adapt the output of the U-Net from the non-target sketch domain to the target sketch domain in few-shot settings. To alleviate overfitting, preserve individual characteristics, and enhance detail representation, we leverage the FFHQ dataset to construct dual training paths and design a domain-directional triple loss. Experiments show that PDGR significantly outperforms previous few-shot learning methods and even outperforms the state-of-the-art FSS methods trained on the full dataset.

Face sketch synthesis, as a key technique for solving face sketch recognition, has made considerable progress in recent years. Due to the difference of modality between face photo and face sketch, traditional exemplar-based methods often lead to missed texture details and deformation while synthesizing sketches. And limited to the local receptive field, Convolutional Neural Networks-based methods cannot deal with the interdependence between features well, which makes the constraint of facial features insufficient; as such, it cannot retain some details in the synthetic image. Moreover, the deeper the network layer is, the more obvious the problems of gradient disappearance and explosion will be, which will lead to instability in the training process. Therefore, in this paper, we propose a multi-scale gradients self-attention residual learning framework for face photo-sketch transformation that embeds a self-attention mechanism in the residual block, making full use of the relationship between features to selectively enhance the characteristics of specific information through self-attention distribution. Simultaneously, residual learning can keep the characteristics of the original features from being destroyed. In addition, the problem of instability in GAN training is alleviated by allowing discriminator to become a function of multi-scale outputs of the generator in the training process. Based on cycle framework, the matching between the target domain image and the source domain image can be constrained while the mapping relationship between the two domains is established so that the tasks of face photo-to-sketch synthesis (FP2S) and face sketch-to-photo synthesis (FS2P) can be achieved simultaneously. Both Image Quality Assessment (IQA) and experiments related to face recognition show that our method can achieve state-of-the-art performance on the public benchmarks, whether using FP2S or FS2P.

Semantic segmentation is a challenging problem in computer vision. Many applications, such as autonomous driving and robot navigation with urban road scene, need accurate and efficient segmentation. Most state-of-the-art methods focus on accuracy, rather than efficiency. In this paper, we propose a more efficient neural network architecture, which has fewer parameters, for semantic segmentation in the urban road scene. An asymmetric encoder-decoder structure based on ResNet is used in our model. In the first stage of encoder, we use continuous factorized block to extract low-level features. Continuous dilated block is applied in the second stage, which ensures that the model has a larger view field, while keeping the model small-scale and shallow. The down sampled features from encoder are up sampled with decoder to the same-size output as the input image and the details refined. Our model can achieve end-to-end and pixel-to-pixel training without pretraining from scratch. The parameters of our model are only 0.2M, 100× less than those of others such as SegNet, etc. Experiments are conducted on five public road scene datasets (CamVid, CityScapes, Gatech, KITTI Road Detection, and KITTI Semantic Segmentation), and the results demonstrate that our model can achieve better performance.