A transfer learning method of collaborating random walk and adaptive instance normalization for inscription image denoising

Mess noise hinders reading and understanding of inscriptions in images. For image restoration from noise-corrupted images, existing network-learning-based methods can construct an excellent model to generate noise patterns. However, the performance of such models is degraded owing to the lack of high-quality training data and the complex noise pattern in inscription images, e.g., mixed noise with multiple levels. Herein, we first propose a novel noise generation model that can produce more realistic synthetic noise images using the random walk algorithm. Then, we propose an explainable inscription image denoising network using a variational inference model, where the joint distribution of clean-noise image pairs is approximated in a dual adversarial manner. The proposed network exhibits improved generalizability and adaptability to different noise characteristics using an estimated noise map and adaptive instance normalization. Finally, we introduce a transfer learning scheme to migrate the network learned from the synthetic noise image domain to a real-inscription image domain with a limited number of real-inscription images. The proposed method outperforms state-of-the-art methods.