Research On Key Technology Of Scene Object And Scene Text Recognition

With the rapid development of economy in the past years,the demand of intelligence deepens industries and daily lives from various aspects.In these techniques,scene image understanding has received more and more attentions due to the fact that it is an essential part of artificial intelligence.In this paper,the recognition of two subjects of scene components are introduced,namely,object detection and text recognition.Both of them have a lot useful applications in real life.For example,when obtaining the class information of a target object,or knowing the meaning of a text in any scene,the source of information can be used by a lot applications and thus the work of people will be greatly reduced.Specifically,we first discuss the background of this research and give an overview of related work.Then we solve the character recognition problem by proposing a new auto encoder based method.Texts from nature scenes face lots of problems,including the variations on fontor color,and complex background.The purposed method enhances particular pixels and weakens the other pixels from background to extract proper feature for scene text detecting.On the training stage,the proposed method not only encodes texts by a global way,but also extracts the information of classes to enhance the ability.When features are obtained,the pooling method is adopted to give more structure information for classifying.Next,this paper introduces a method to detect in-door objects based on Hough Forest.This method uses random forest,and adds location information to nodes besides object class information.By this way,the recognition performance of scene objects can be improved.Experimental results for character recognition and in-door objects detection on the benchmark character dataset and our object dataset show the proposed method is superior to the other methods in the aspects of recall,precision,and speed.