Research Of Character Detection And Recognition In Natural Environment Based On Neural Network

In recent years,with the popularity of mobile networks and the massive use of mobile phones,people have more and more ways to obtain and share pictures from natural scenes,and it has become more and more important to recognize the characters contained in massive natural scene pictures.Smart cities,industrial automation,etc.all need to recognize scene characters in real time.Scene character recognition belongs to the scope of text recognition.In traditional recognition methods,it is often necessary to design features manually.In view of the diversity and randomness of the background,this traditional algorithm often cannot meet the actual needs.The rapid development of deep learning technology has brought new opportunities for character detection and recognition in natural scenes.Compared with traditional recognition methods,neural networks can not only automatically extract image features through convolution,but also avoid the huge workload of manually designing features in traditional methods.Therefore,the use of neural networks for character recognition of natural scenes has become the main research direction right now.Although the current character recognition algorithms in natural scenes have made great progress,there are still the following problems: the detection of small targets is not good;it is difficult to distinguish sticky text;for curved text,the detection and recognition effects have always been not good.In response to the above problems,this article made the following innovative improvements,the specific research work is as follows:1.For the problem of small target detection effect is not good,this article combines the detection method based on pixel segmentation and the erosion and expansion algorithm in morphology,remove the unrelated small target through corrosion,and then expand to expand the real small target area,This can effectively detect small targets and increase the accuracy rate from 81.8% to 85.1% on the Total-Text data set.2.For the text that is sticky,this paper combines the area expansion algorithm and deformable convolution,and uses deformable convolution to extract features from multiple angles,thereby improving the detection efficiency of the sticky text.3.For the detection and recognition of curved text,this paper combines multi-point marking of text area and area expansion algorithm during detection,which has a good detection rate for curved text.The accuracy rate of detection on CTW1500 data set is from 80.6% Increased to 83.2%.4.The CRNN and Attention mechanisms are combined during recognition to improve the recognition rate of the text.