Image Correction And Text Recognition For Deformation Label Of Storage Package

The digitization of paper documents is of great significance for information extraction,text content analysis and cultural dissemination.This thesis focuses on the digitization of labels on logistics storage packages,which belongs to the digitization of paper documents in special scenarios.In the process of transportation and storage,the labels on the surface of the storage packages suffer from geometric deformations such as folding,curled and wrinkles.At the same time,collectors often hold mobile devices to capture label images,which brings problems such as irregular collection angles,uneven illumination conditions and blurred images,making it difficult for common scene text detection and recognition methods to extract text information in label images.In order to solve the above problems,this thesis focuses on image correction and recognition of deformation label on storage package.The main research contents are as follows:(1)Aiming at the problem of the lack of public deformed document image datasets,this thesis constructs a complex deformed label image dataset.The training set of this dataset uses the depth camera Kinect V2 combined with the Elastic Fusion algorithm to capture the three-dimensional shape of the deformed label in the natural scene.Rendering in the 3D rendering software Blender generates a large number of deformed label images and their corresponding rich annotation information,which are used for the training of the image correction network.After that,a small part of the synthetic images in the training set and the real deformed label images in the test set are annotated with the text area coordinates and text transcription,which are used for the training and testing of the text detection and recognition network.(2)Aiming at the problem of image correction of complex deformed labels,this thesis studies the existing traditional image correction methods and image correction networks based on deep learning.Finally this thesis adopts a combination of traditional image correction preprocessing based on label extraction and affine transformation and Dewarp Net network image correction method.On the basis of the original Dewarp Net network,this thesis replaces the preprocessing steps of cropping the input image required by the original network with a method based on label extraction and affine transformation,which automatically preprocesses the input image;in the loss function of the Dewarp Net texture mapping subnetwork introduces the SSIM loss item.Finally,through a large number of experiments,it is verified that compared with the original network,the improved Dewarp Net network has better image correction performance the multi-scale structure similarity in the data set is improved by 0.0252 and the local distortion is reduced by 1.3777.(3)Aiming at the problem of text detection and recognition of the corrected label image,this thesis studies a text detection method that combines the Text-context-aware Attention Module and Mask Text Spotter v3 network.After that,the detected polygonal text area uses minimum circumscribed rectangle and generates a mask for the text area.Finally,the detected text area is sent to the CRNN+CTC network for text recognition.The experimental results show that the method can accurately detect multi-oriented,extreme aspect ratio and distorted text lines and effectively suppress situation of over-segmenting long text lines and generating false positive detected result.Compared with the original network,the improved text detection network improves the accuracy by 1.97% and F-measure by 0.91%.