Research On Scene Text Detection Method Based On EAST Algorithm

In the field of deep learning,scene text detection is usually regarded as a special form of object detection task.Object detection is a computer vision technique used to detect different objects in an image.Compared with universal object detection,scene text detection has some special challenges.The shape and size of the local area in scene text are inconsistent,and the direction and tilt Angle of the text may be different.In addition,the text area is usually smaller than other objects,and may even be very fuzzy,which means that the algorithm needs to be optimized.Given the above situation,scene text detection is still a highly challenging task to this day.This thesis takes EAST scene text detection algorithm as the framework for research,and the main research contents are as follows:(1)An improved EAST algorithm based on strip pooling is proposed.Firstly,in the EAST algorithm,stripe pooling module(MPM)is introduced into the feature fusion layer,which effectively expands the receptive field.Strip pooling is a long and narrow pooling core,which can better capture the local information in the long and narrow shape of long text,so that the pixel belonging to the text can contain the information at the farthest text boundary.Secondly,the Coordinate Attention mechanism is introduced to convert position information into attention weights by capturing orientation perception and position sensitive information,and these weights are used to calculate the weighted sum of feature maps to help locate long text accurately.The proposed algorithm was tested on the ICDAR2013 and ICDAR2015 data sets,and compared with the EAST algorithm,the ability of long text detection was improved.(2)An improved EAST algorithm based on SENet and mixed loss function is proposed.Firstly,SENet attention mechanism is introduced into the feature fusion layer in the EAST algorithm to assign greater weight to the shallow feature map channels with detailed features,so as to emphasize the detailed features in the shallow network and improve the effect of short text detection.Secondly,the model is optimized by mixing loss function and CIo U loss function to further improve the detection performance of short text and even all text.The hybrid loss function uses Dice loss function and quasi-equilibrium cross entropy loss function to combine their advantages and complement each other.This algorithm was tested on the ICDAR2013 and ICDAR2015 datasets and improved its ability to detect short text compared to the original algorithm.(3)An improved scheme of scene text detection algorithm based on Res Ne St network is proposed.First,the Res Ne St-50 network is used in the feature extraction layer of the EAST algorithm.By introducing Split-Attention Networks in the Res Ne St-50 network,the feature mapping attention can be carried out between different feature mapping groups,so that the model can learn more feature correlation.Thus improve the efficiency and accuracy of feature extraction.The algorithm was tested on the ICDAR2013 and ICDAR2015 data sets,and compared with the original algorithm,the ability of feature extraction has been improved to some extent.Secondly,the Res Ne St-50 network extension is applied to the Advanced EAST algorithm,and the experiment is carried out on the Tianchi 2018 data set.Compared with the original algorithm,the feature extraction ability has been improved to some extent.