Label Text Recognition Methods In Wild Images

Label text is used to display product information or identify product instances,which is generally composed of numbers,letters,Chinese characters,and special characters.Compared with ordinary text,label text has no obvious semantic information and grammar specification,and thus cannot be recognized using contextual information for correction.In addition,ordinary text recognition scenes are standardized and controlled,while label text recognition scenes such as license plate and terminal block are challenged by uneven illumination,variable shape and complex background.In particular,this thesis selects license plate recognition and terminal block label recognition for two reasons:First,license plate and terminal block recognition have important application value,the former facilitates traffic management and maintains traffic safety,the latter guarantees the correct wiring of electric power equipment and maintains the safety of substations.Second,the current license plate and terminal block recognition methods requires high image quality,and the recognition performance is poor for cases such as incorrect imaging angle and blur.Specifically,this thesis proposes a method based on the holistic position attention(HPA)for license plate recognition that explicitly extracts the position information of characters;and proposes an overall scheme containing detection,recognition,classification and matching for terminal block recognition.The main contributions of this thesis include:(1)A novel holistic position attention,consisting of a position network and a shared classifier,which is designed for license plate recognition.The position network explicitly encodes the character position information into the maps of HPA,and the shared classifier performs the character recognition in a unified and parallel way.Here the extracted features are modulated by the attention maps before feeding into the classifier to yield the final recognition results.Our proposed method is end-to-end trainable,character recognition can be concurrently performed,and no post-processing is needed.Experimental results on several datasets demonstrate that the method can achieve good effectiveness and efficiency simultaneously,and is able to recognize multiple types of license plates.(2)A terminal block recognition scheme based on convolutional network,which is designed according to the requirements of terminal block recognition.Firstly,for terminal block text detection,Improved_EAST model is proposed,which introduces atrous convolution and PPM module on EAST network to enhance the capability to predict long text regions.Then,image correction is introduced before terminal block text recognition to make the image easy to recognize.Finally,text classification and Kuhn-Munkres algorithm are used to match the serial number with the number pipe to achieve the final purpose of terminal block recognition.The results of qualitative and quantitative analysis demonstrate that the proposed scheme has good robustness.In summary,this thesis makes an in-depth exploration of license plate recognition and terminal block recognition using the characteristics of label text.The proposed license plate recognition method can achieve high recognition accuracy and computational efficiency simultaneously.The proposed terminal block recognition scheme has good performance in various indicators,and the research results have practical application value.