Feature Selection Based Multi-oriented Scene Text Detection

In recent years,the wide application of deep learning has provided a new impetus for the development of text detection tasks in natural scenes.In order to further improve the performance of text detection methods,this paper proposes a Location-aware Feature Selection Text Detection Network and Performance-Guided Multi-stages Features Network for Scene Text Detection to improve the performance of text detection from the perspective of feature utilization.In common natural scene datasets,the method in this paper can achieve state of the art performance.Among the text detection methods,the text detection method based on direct regression has attracted the attention of many researchers because of its simple network structure and stable detection performance.This method is lacking in detection accuracy,especially for the detection of long text and large text,it is difficult to accurately predict the text boundary.Therefore,in the scene that requires accurate text detection results,its practical value has been greatly affected.To solve this problem,this paper proposes a novel location-based feature selection network.The reason why it is difficult to obtain accurate text boundaries is that it directly predicts a complete text box only by using the features of a single location.This method not only limits the utilization efficiency of features,but also does not take into account the different requirements of different components of the text box for features.The method proposed in this paper disassembles the multi-directional text box,uses the features of a single position to predict the box components respectively,and then combines the accurately predicted components to get the final detection box.This method greatly improves the utilization of features.At the same time,considering the different requirements of different components for features,more accurate components are predicted,so as to obtain more accurate text prediction boxes.With the improvement of hardware computing power and the progress of optimization methods,deep neural networks can be effectively trained.The features at different stages of the deep neural network contain different feature information.In the text detection task,the diversity of text objects has different requirements for features.Therefore,how to reasonably use the features of different stages of the network has become a breakthrough to improve the performance of the text detection task.In the existing work,most of the work achieves the matching of detection targets and features in different stages through prior knowledge,so that the features in different stages can be trained at different detection targets.However,it is not reliable to match features and detection targets through prior knowledge.The main reason is that it is difficult to construct appropriate prior information to comprehensively and accurately express the adaptability of features and targets.This paper proposes a performance-guided task assignment strategy.In this strategy,we can judge the adaptability of the detection target in each stage according to its performance,and then control the influence of the detection target on each stage feature training according to the obtained adaptability factor.Experiments show that method can greatly improve the performance of the network in text detection tasks.