Research On Key Technologies Of Character Recognition In Natural Image

Characters carry more information than other contents in a natural image.Character recognition play an important role in image retrieval,image analysis and image understanding,Therefore it receives more and more attentions nowadays.The traditional character recognition technology(OCR)has been developed for many years.However,it only can be used for recognizing the image of scanned document with relative clean background,high resolution and high contrast.However,the performance deteriorates when we use OCR to identify characters in the image with cluttered background.This is because there are many problems such as lighting conditions,perspective transforms,cluttered background,etc.Thus,the main objective in this paper is how to overcome these obstacles in character recognition process.In this paper,we most focus on two key technologies in character recognition,feature extraction and feature selection.Details are as follows:(1)We proposed a method of characters recognition in nature image by adopting integral channel features and pooling strategy.Base on the existing research,the method of image feature expression improved combined with the pooling strategy,the multi-channel characteristics and related parameters verified by experiments.Experimental results show that our proposed method has strong adaptability to text recognition in natural scene images.Compared with other methods,the proposed method has the advantages of simple structure,high computing efficiency and high accuracy.(2)Natural images have a large number of irrelevant and redundant features due to its complicated background.To solve this problem we proposed a Multiclass feature selection algorithm based on Relief F and MSVM-RFE.We mainly improved the feature evaluation function,used the weights of Relief F and MSVM-RFE two algorithms to evaluate the image features,and increased the weight ratio of the MSVM-RFE algorithm in the iterative process.In the experiment,we compared our algorithm with several multiclass feature selection algorithms.The results show that our proposed Relief F-MSVM-RFE algorithm can achieve much better performance in the practical classification task,and has a great value of practical application.