Scene Text Detection And Recognition

Text in natural scene images contains rich high-level semantic information which is crucial for scene image understanding.Due to the diversity of natural scene,scene text detection and recognition is full of challenges.Although its performance has been significantly improved compared to traditional methods under the promotion of deep learning,there still are many problems that need to be solved.Scene text understanding is a hot topic in the field of computer vision.There are four main works:(1)Scene text detection is considered as a dense regression task,which combines non-local neural networks to embed global information to expand the receptive field,making it better applicable to multi-oriented text.The validity is verified on a multi-oriented scene text benchmark.(2)From the perspective of instance-aware segmentation,a feature fused text instance segmentation network is proposed for text detection.This network can detect text with arbitrary shapes efficiently and accurately.At the same time,based on the characteristic of text,the nonmaximum suppression post-processing method is improved which achieves superior performance on both multi-oriented and curved text datasets.(3)A deep-supervised scene text recognition algorithm combined with attention mechanism is proposed which is improved on the current mainstream text recognition framework.Without introducing parameters and computations at the inference stage,one extra supervision module is added during the training process which allows the model to get better feature representations.At the same time,the attention mechanism is further integrated to make the model adaptively suppress redundant clutter and focus on important text areas.The effectiveness of this method is verified on several text recognition benchmarks.(4)An end-to-end scene text recognition network is put forward which can recognize text in any shape,integrating detection and recognition in the same network,allowing these two tasks to share some parameters,reducing the overall parameter and computation of the model.At the same time,these two tasks can benefit from the supervision signal of each other,making the entire model more robust.Experiments are conducted on horizontal and curved text datasets,and the results are in line with expectations.