Research On Degraded Document Image Binarization Methods Based On Fully Convolutional Networks

Degraded document image binarization is a research hotspot in the field of document analysis and recognition.Because documents are subject to physical conditions during storage,leading to paper breakage,ink stains,background penetration,or smudging due to human damage,and uneven illumination when shooting document images,result in a large degree of document images.Degradation seriously affects the effect of algorithm processing such as document analysis and character recognition.Therefore,degraded document image binarization has extensive research and application value.The main research contents and work of this thesis are as follows:(1)To solve the problem that the thickness of the strokes is different and the weak illumination causes poor binarization after global sharpening,a degraded document image preprocessing method with local adaptive sharpening combined with illumination compensation is proposed for degraded document images.In the experiment,it is found that the degree of sharpening has different effects on the strokes of different thicknesses.As the degree of sharpening increases,the fine strokes retain more information in the binarization results,while the thick strokes gradually appear to be broken.Aiming at this problem,the stroke width estimation algorithm is used to estimate the stroke width of the local area,and adaptive sharpening is performed to solve the problem of cracking of the thick stroke while retaining more fine stroke information.In addition,sharpening effects on intensely illuminated document images are more pronounced than weakly illuminated document images,resulting in better binarization results.Therefore,this thesis makes a suitable illumination compensation for the image of the weak illumination document as a useful complement to image sharpening.The experimental results show that after the degraded document image is processed by the proposed pre-processing method,the binarization result restores more detailed information,while retaining the coarse stroke information,and the binarization effect is improved.(2)For the degraded document image binarization,the data set is small,making the network unable to be fully trained or easily over-fitting,and the single-convolution network has poor generalization ability,resulting in poor binarization performance.A degraded document image binarization method based on migration learning and full convolutional network is proposed.U-Net,a better-convolutional network model that performs better on smaller data sets,is selected and migration learning techniques are introduced.Three commonly used migration learning models of VGGNet,ResNet and Inception are selected as pre-training encoders for U-Net networks;different decoder network structures are established for different model characteristics;and fusion with U-Net networks,which is different from the conventional processing of the output of the pooling layer,the output of the convolutional layer is used as a jump connection and fused with up-sampling,so that the up-sampling layer can better restore the detailed information of the document image.The experimental results show that the U-Net network with migration learning model accelerates the convergence speed and improves the generalization ability of the model,thus improving the binarization effect of degraded document images.The binarization performance of the binarization method including the degraded document images of VGG16 and U-Net(denoted as V16_U-Net binarization method)is more prominent.(3)A comprehensive experiment was carried out on the DIBCO and Palm Leaf Manuscripts degraded document image datasets using the proposed local adaptive sharpening combined with the image compensation preprocessing method of illumination compensation(denoted as LASIC)and the V16_U-Net binarization method.The experimental results show that the proposed V16_U-Net binarization method has achieved good results on multiple data sets,and the final binarization results combined with the LASIC preprocessing method and the V16_U-Net binarization method have been further improved in various evaluation indicators,indicating that it has better performance in degraded document image binarization.