Character Recognition Based On Neural Network For Indoor Scenes

The traditional character recognition recognizes image from the scanner,requiring high quality images to achieve better recognition results.Nowadays,with the rapid development of image processing technology and artificial intelligence,as well as the popularity of intelligent devices with shooting capabilities,the method of obtaining character information in images is no longer limited to scanners or manual input methods,while turning to these mobile devices with camera functions.The indoor scene is the main environment of people`s life,occupies the great part of people's living environment.To get these huge character information effectively will greatly help people understand the content of the scene,and provide convenience for people.We have studied character recognition in indoor scenes.First,we summarize the research status of character recognition at home and abroad.Then,we introduce the character detection,character segmentation and character recognition.The main research works are as follows:(1)Based on improved stroke width transformation,this thesis proposes a character detection algorithm by using the maximum stable extremal region region(MSER)algorithm and the stroke width transformation(SWT)algorithm.Firstly,we use multi size of image and multi threshold MSER to detect the character regions.Then we use the SWT algorithm to compute the character stroke width of the candidate regions.Finally we use classifier to remove the non text regions.(2)This thesis propose an improved PCANet network structure by analyzing the PCANet network.(3)The network is improved on the classical LeNet-5 network structure,which makes the network converge when training little data.By comparing the performance of the improved network and the improved network,we can conclude that the improved network has better performance.In addition,this thesis also studies and analyzes several aspects that affect the PCANet network.We collect five kinds of images from indoor scenes.Then we build a small data set for character detection and a set of implementation process from character detection to character recognition.In this thesis,the improved text detection algorithm is compared with the SWT algorithm in the detection phase.The experimental results indicate that the improved algorithm in the detection effect has been greatly improved.In the recognition stage,this thesis use four methods to recognize the character and analyzes the experimental results.