Research On Text Location Method In Natural Scene And Application In Intelligent Device

Text recognition in nature scene is currently applied in various intelligence equipment.As a crucial interface between the equipment and the natural environment,an excellent text recognition algorithm that can produce accurate results is pursuit.In the process of text recognition,the first job is to locate the text precisely.The proposal-based methods are efficient for horizontal and oblique texts;however,they usually have a poor performance when being applied to curved texts.Curved text is very different from horizontal and oblique text in terms of text character connection and marking criteria,which makes the design of candidate box and bounding box regression more difficult.Segmentation-based method combined with the adaptive text box marking method can solve this problem well.Firstly,the text area is accurately segmented by the segmentation-based method,and then accurately positioned by the adaptive text box method,which can effectively detect and locate complex curved text.In this thesis,the text detection and location in natural scene is studied based on the learning method of pixel segmentation and adaptive text box calibration.Finally,the method is applied to the intelligent device platform.(1)An improved text detection method is proposed.It incorporates the attention mechanism into the current model to improve the detection performance of the text recognition in natural scenes.To overcome the issue of poor overall performance in current Pixe Link algorithm,the weights are imposed on the feature channels.By improving the weight coefficient of effective feature channels and suppressing the ones of inefficient or invalid channels,an improvement is achieved on the robustness of overall text detection and localization eventually.Experimental results on ICDAR2015 and MSRA-RD500 datasets show that the proposed method can improve the overall text detection accuracy by 1.2% and 3.5% respectively compared with the original pixellink method.Therefore,the proposed method can effectively introduce feature channel attention mechanism to improve the detection performance of the original method.(2)A method to change the annotation form of public dataset and adaptive text box location is proposed.By changing the form of dataset annotation,the coordinate points can be expressed as a series of sequence forms,so that the text lines can be framed adaptively in the LSTM model.At last,the located object is rotated according to the angle between a pair of vertexes in the polygon frame,and is subsequently fed to the text recognition interface to obtain the final character.Compared with textsnake,the proposed method is 12.1% and 2.6% better than textsnake in terms of accuracy and comprehensive index.Therefore,the proposed method can locate curved text effectively and achieve a leading position.(3)An interactive reading guide robot dog is developed.The design includes a set of complete hardware structure combination scheme and function realization scheme,after combination and installation,the user can interact with the information through the voice interface and navigation interface,so as to increase the communication channel between the blind and the outside world.The text detection method is applied to the robot dog,and the identification test is carried out in outdoor and indoor scenes to complete the function of interactive reading accompaniment.Besides,some efforts are also put on the designing and verification of other fundamental functions of the robot dog,which improves the driving security,and completes a fully functional equipment platform eventually.