Image restoration is a prominent field of research in computer vision. Restoring broken paintings, especially ancient Chinese artworks, is a significant challenge for current restoration models. The difficulty lies in realistically reinstating the intricate and delicate textures inherent in the original pieces. This process requires preserving the unique style and artistic characteristics of the ancient Chinese paintings. To enhance the effectiveness of restoring and preserving traditional Chinese paintings, this paper presents a framework called Sketch-Guided Restoration Generative Adversarial Network, termd SGRGAN. The framework employs sketch images as structural priors, providing essential information for the restoration process. Additionally, a novel Focal block is proposed to enhance the fusion and interaction of textural and structural elements. It is noteworthy that a BiSCCFormer block, incorporating a Bi-level routing attention mechanism, is devised to comprehensively grasp the structural and semantic details of the image, including its contours and layout. Extensive experiments and ablation studies on MaskCLP and Mural datasets demonstrate the superiority of the proposed method over previous state-of-the-art methods. Specifically, the model demonstrates outstanding visual fidelity, particularly in the restoration of landscape paintings. This further underscores its efficacy and universality in the realm of cultural heritage preservation and restoration.

Calligraphy (the special art of drawing characters with a brush specially made by the Chinese) is an integral part of Chinese culture, and detecting Chinese calligraphy characters is highly significant. At present, there are still some challenges in the detection of ancient calligraphy. In this paper, we are interested in the calligraphy character detection problem focusing on the calligraphy character boundary. We chose High-Resolution Net (HRNet) as the calligraphy character feature extraction backbone network to learn reliable high-resolution representations. Then, we used the scale prediction branch and the spatial information prediction branch to detect the calligraphy character region and categorize the calligraphy character and its boundaries. We used the channel attention mechanism and the feature fusion method to improve the detection effectiveness in this process. Finally, we pre-trained with a self-generated calligraphy database and fine-tuned with a real calligraphy database. We set up two groups of ablation studies for comparison, and the comparison results proved the superiority of our method. This paper found that the classification of characters and character boundaries has a certain auxiliary effect on single character detection.

Line drawing, characterized by its unique techniques and profound cultural heritage, presents the boundless charm of traditional Chinese art. To gain deeper insights into the esthetic styles of different dynasties and the essence of various painting schools, employing digital methods to extract the contours of relics is essential. Although existing digital line drawing extraction methods can enhance efficiency, the surfaces of relics often exhibit pigment mottling, cracks, and other diseases. These interfering elements tend to intertwine with the contours, increasing the difficulty of line extraction. For this task, we propose a knowledge-sharing strategy-based approach to learn the mapping from relic images to contours in a semi-supervised manner. We integrate a supervised learning branch into the unsupervised image translation framework that utilizes undamaged relic images paired with corresponding noise-free contours for training. By sharing the generator, the noise-free contour style learned by the supervised branch is infused into the unsupervised branch, effectively suppressing diseases. Furthermore, we achieve complementary texture features through two independent streams: the contour generation flow (CGF) and the auxiliary gradient flow (AGF), each designed with separate GAN architectures. To better capture relic details, we introduce the coordinate attention transfer (CAT) module, which adaptively transfers features from AGF to the corresponding dimensions of CGF by learning spatial orientation and positional weights. Finally, the bidirectional gated transposed fusion (Bi-GTF) module employs two soft gates to regulate the information fusion between CGF and AGF, guiding CGF to focus on the shape and intricate details of the relics. In our experiments, we conducted both quantitative and qualitative comparisons with six state-of-the-art algorithms, analyzing the impact of gradient-guided feature transfer and the generalization of the semi-supervised framework. The results demonstrate that Relic2Contour generates rich, clear, and high-quality contour images from relics across various styles.