Enhancing Scene Text Segmentation through Subtask Decomposition
Yong Wang, Youguang Chen · 2023
The field of image processing widely utilizes scene text segmentation technology, with applications extending to image editing and font style transfer. These applications enhance image understanding quality and aid in boosting the performance of numerous computer vision tasks. The advent and progression of deep learning have led to substantial advancements in scene text segmentation technology. However, the limited size of existing scene text segmentation datasets constrains the performance of models. Therefore, we propose an algorithm for synthetic segmentation data. We first pretrain the model using large-scale synthetic data, then fine-tune it on the target dataset to address the issue of limited dataset size. Existing models employ end-to-end segmentation, which presents challenges in segmentation. We propose a scene text segmentation method. By decomposing the segmentation task into subtasks and solving them one by one, the complexity of the task can be reduced compared to direct segmentation of the entire image significantly improving the segmentation effect. The proposed method consists of three modules: a fragment crop module, a fragment segmentation module, and a fragment combination module. The fragment crop module is composed of an additional corp layer added after DBnet. The fragment segmentation module can be embedded with various segmentation methods. The fragment combination module uses the maximum pixel value pasting algorithm to combine the segmented fragments. We call this method Crop-Segmentation-Combination Framework (CSCF). We conducted experiments on the ICDAR 2013 and TextSeg datasets. The CSCF, embedded in Unet within the segment segmentation module, enhanced the text segmentation IoU by 5.80% on the ICDAR 2013 test dataset. Our suggested approach has been shown to notably enhance the efficiency of scene text segmentation.