Scene Text Erasing Algorithm Based On Multi-Branch Network Fusion

Various text messages often appear in natural scenes.The presence of these text messages can sometimes cause some trouble in the use of the pictures.For this reason,there is a pressing need for a method that can effectively remove text from images while preserving and restoring the original background information at the text as much as possible.For the above-mentioned scene text erasure task,this paper proposes two scene text erasure algorithms based on multi-branch fusion network,which effectively solve the problem of text omission and over-erasure that often occurs in the scene text erasure task.The main work and innovation points of this paper are as follows:1.multiple generator networks and discriminator networks based on scene text features are constructed to realize multiple erasures of scene text,and multiple branch networks are fused by multi-level feature fusion mechanism to finally form a baseline system for scene text erasure.In this paper,this approach is called multi-branch fusion GAN-based scene text erasure algorithm.2.In order to achieve end-to-end high-quality scene text erasure,an algorithm based on complementary fusion of multi-branch networks is proposed.Our algorithm decomposes the scene text erasure task into two parts:text erasure and background retention,and accomplishes these two tasks by introducing multiple complementary branching networks and the fusion processing between them.The experimental results verify the effectiveness of the proposed algorithm.3.To further improve the effect of text erasure,an information processing scheme is proposed to fuse multiple branched networks simultaneously from both feature and image levels,so that the fusion results can take the strengths and weaknesses from each branched network and finally generate the desired erasure results.4.To validate the performance of the proposed multi-branch complementary fusion network scene text erasure algorithm,relevant performance tests are conducted on two publicly available datasets.The experimental results show that our algorithm performs well on the public dataset and achieves the best performance metrics so far.