Scene handwritten text erasure based on multi-scale feature fusion

The removal of handwritten text from diverse scenes is increasingly important for various applications such as privacy protection, data cleaning, and image repair. However, the current focus of deep neural networks is primarily on printed text erasure, with limited research on erasing handwritten text. Furthermore, existing methods struggle to effectively remove handwritten text in complex scenes. To tackle this challenge, this paper proposes a two-stage network denoted multi-scale feature fusion network (MSFF-Net). In the initial stage, a multi-scale feature fusion attention mechanism (MSFF-AM) is proposed to accurately identify areas containing handwritten text. In the second stage, incompletely erased handwritten text is treated as image noise and addressed with an iterative data refinement network. To handle the challenges of image restoration, a multi-scale feature extraction (MSFE) module is designed to capture image features effectively, while combining a non-local module to maintain the integrity and consistency of the final restored image by establishing long-range dependencies. Moreover, in support of research efforts on scene handwritten text erasure, a large-scale dataset for scene handwritten text removal (HDU-SHTR) is created for this specific purpose. Experimental results on scene handwritten text datasets demonstrate that MSFF-Net significantly outperforms existing methods and exhibits excellent performance on other public text erasure datasets.