Research On Historical Document Image Binarization

Document analysis and recognition includes image preprocessing,text feature extraction and database comparison,binary segmentation precision plays an important role in the whole system and the accuracy of binary segmentation directly affects the reliability of recognition results.Degraded historical document images have various degradation factors,Such as page stain,character fading,fine strokes,uneven illumination and ink bleed through and folding mark.As a result,the existing algorithms are difficult to achieve better segmentation results.Therefore,this paper focuses on the binarization algorithm of degraded historical document images.The main contents and innovations are as follows:(1)To solve the problem of low contrast of degraded document image,this paper present a novel method based on background estimation and energy function.Firstly,the min-average method is adopted as gray preprocessing,which reduce the variation intensity within text class and enhance the contrast of the image.Secondly,the stroke width is obtained based on the entropy of stroke information,and estimates the document background via morphological closing operations.Finally,the image is mapped to the energy function,and the graph structure based on the energy function is constructed.The optimal solution of the energy function is obtained by the minimum cut of the graph.The experimental results show that the proposed algorithm can effectively deal with all kinds of degraded document images and suppress the background of stains.(2)In view of a noise problem in dealing with the binary results of degraded historical document images by the above algorithm,This paper put forward a binarization algorithm based on Hybrid Pyramid U-shape convolution neural network(HPU-Net).Firstly,the background estimation process of the above algorithm is used to obtain the enhanced document image after background removal.then the logarithmic loss and Dice coefficient are combined as loss function,and a convolutional network named HPU-Net is adopted to classified image pixels from the foreground document.Finally,the image is finely binarized by Otsu algorithm.The experimental results show that the binarization algorithm based on HPU-Net can not only preserve the stroke details,but also achieves better segmentation results.(3)The two algorithms proposed in this paper are compared with those of the thirteen binarization algorithms by using the data set and evaluation indicators of the Document Image Binarization Contest(DIBCO).Experimental results show that our proposed method outperforms other state-of-the-art document image binarization algorithms in terms of F-Measure(FM),pseudo F-Measure(p-FM),Peak Signal to Noise Ratio(PSNR)and Distance Reciprocal Distortion(DRD),with 5.00%?5.82%?1.22 dB and 2.86%.