Research On Natural Scene Text Detection Algorithms Based On Deep Learning

Intelligent robots' perception of the external environment depends on the understanding of the scene.Text is an indispensable part of the scene content.Text detection in natural images has become an important research direction in the field of intelligent robots.The Faster R-CNN algorithm has advantages in accuracy and speed for object detection.However,it can not detect text in natural images as well as other objects because of the variability of text and the interference from external factors.Moreover,the existent text detection algorithms based on deep learning need large data sets to train the network.In some special cases,it is impossible to obtain a large amount of training data,and the performance of this kind of algorithms will be limited.It's an engineering problem that how to detect text in natural scenes based on small data sets.To address these issues,this paper researched on text detection in natural images based on deep learning with small data sets.A multi-scale text feature extraction network based on feature pyramid was proposed,which can accurately and comprehensively express complex and changeable text features in natural scenes even though the training samples were very small.The structure of feature pyramid can fuse the semantic information of deep features with the location information of shallow features.In this way,the influence on the test results caused by the diversity of text was overcome,the small text was not lost in the detection process,and the accuracy of text detection was improved.Faster R-CNN was modified to adapt to the particularity of the text.The multi-scale feature extraction network was used to extract text features.The size and proportion of the initial anchors were modified to ensure that the candidate boxes were conformed to the size rule of the text.The training method of the model was researched according to the characteristics of the text,and the appropriate number of initial candidate boxes and the threshold for classification of foreground and background were explored.The algorithm was implemented by TensorFlow.The network was trained on a large scale data set which contains a small number of training samples.The model was evaluated by precision and recall.Experimental results demonstrate that the proposed method outperforms conventional Faster R-CNN and is very competitive with existing architectures that have a similar framework.This paper explored an effective solution to natural scene text detection,and the results have engineering application value.