MGPAN: MASK GUIDED PIXEL AGGREGATION NETWORK

Recently, semantic segmentation and instance segmenta-tion methods have been widely adopted by arbitrary shaped text detection tasks. In this paper, we propose a combined framework termed Mask Guided Pixel Aggregation Network (MGPAN), which takes advantage of the accuracy of the segmentation method and of the efficiency of the anchor-based method. Specifically, the segmentation branch learns texts, kernels and similarity vectors, and the mask branch learns bounding boxes and masks of texts just like one stage instance segmentation tasks. Both of these two branches share the backbone and the feature pyramid enhancement modules. Then, a kernel enhancement algorithm is applied to accelerate the post processing process. It reduces time cost of post processing, which is helpful in batch processing pictures. Quantitative measurements validate the efficiency of the proposed method in detecting arbitrary shaped text.