A Lightweight NUet Model for Document Image Restoration

The automatic analysis of document images has important applications. However, the quality of document images often cannot meet the requirements, as device aging, improper operation or environmental factors can cause document images to contain noise and blur. Existing document image restoration techniques have high computational complexity. To address this issue, firstly, we proposed three new convolution modules: BSCM model, spindle shape module and hourglass shape module, which are all combining the spatial separable convolutions with the depthwise separable convolution. Then, we propose a lightweight document image restoration model (NUnet), which combines U-shape as backbone network and apply the proposed three new convolution modules to build the network. Finally, we evaluated the proposed model in Kaggle and GOPRO dataset. Experiments have confirmed that the proposed network model has a small number of parameters, and the visual restoration results are superior to other models.