Document Binarization Using Topological Clustering Guided Laplacian Energy Segmentation

Abstract

Abstract-The current approach for text binarization proposes a clustering algorithm as a preprocessing stage to an energy-based segmentation method. It uses a clustering algorithm to obtain a coarse estimate of the background (BG) and foreground (FG) pixels. These estimates are used as a prior for the source and sink points of a graph cut implementation, which is used to efficiently find the minimum energy solution of an objective function to separate the BG and FG. The binary image thus obtained is used to refine the edge map that guides the graph cut algorithm. A final binary image is obtained by once again performing the graph cut guided by the refined edges on a Laplacian of the image.

Historical Document Image Binarization: A Review

“…dev. AdOtsu [95] Local threshold Local version of Otsu Lu [82] Edge based Local thresholding near edges after background removal Su [162] Edge based Filters canny edges using local contrast Jia [53] Edge based Detecting symmetry of stroke edges Valizadeh [166] Edge based Adaptive water flow model Hadjadj [43] Edge based Active contours initialized using contrast edges [162] Rivest [139] Edge based Level-set method Nafchi [104] Image transform Threshold filter response in frequency domain Sehad [147] Image transform Performs background removal using Fourier Transform Zemouri [177] Image transform Uses Contourlet Transform to smooth image FAIR [73] Mixture model Ensemble of MoG with post-filtering Hedjam [47] Mixture model MoG with spatially varying Σ k Mishra [91] Mixture model 10-component MoG for foreground color variation Mitainoudis [93] Mixture model MoG over pairs of intensity co-occurrences Ramirez [138] Mixture model Mixture of log-normal distributions outperforms MoG Howe [50] CRF Laplacian unary term and pairwise Canny-based term Ayyalasomayajula [7] CRF Pairwise terms based on an initial binarization Peng [120] CRF Pairwise terms based on an initial foreground skeleton Su [161] CRF Uses CRF to classify uncertain pixels Ahmadi [2] CRF Learns linear combination of feature functions GiB [13] Game theory Extracts features for clustering using game theory Hamza [44] Shallow ML Self-organizing map to cluster pixels Rabelo [132] Shallow ML MLP to classify pixels using local mean Kefali [58] Shallow ML MLP using local intensities and global statistic features Pastor [118] Shallow ML MLP with F-measure loss function Kasmin [56] Shallow ML Ensemble of 8 SVMs Wu [174] Shallow ML Random forest trained on a rich feature set Pastor [119] Deep learning First CNN for binarization Peng [121] Deep learning Encoder-decoder FCN trained on synthetic data Calvo-Zaragoza ...…”

Historical Document Binarization Combining Semantic Labeling and Graph Cuts

“…This theoretical framework helps to define a topology [1] within this space with a natural way of defining hierarchical clusters with neighborhood constraints. Fig.2 shows the space S for a typical image from the DIBCO dataset with points in red and blue representing the FG and BG pixels, respectively.…”

“…Although the fundamental idea governing Howe's method have been explored previously as separate methods, combining them into an energy function proved particularly effective. Further improvement of Howe's approach was proposed by Ayyalasomayajula and Brun [1] through detection of seeds for the source and sink estimate with effective detection of edges by exploiting the inherent topology by defining a binarization space.…”

“…The proposed method improves upon our previous work in Ayyalasomayajula and Brun [1]. The method builds upon three ideas, basic idea is borrowed from the Howe's approach [4] of defining the binarization as labeling problem.…”

“…The method builds upon three ideas, basic idea is borrowed from the Howe's approach [4] of defining the binarization as labeling problem. The next step is to incorporate the benefits of defining good seed points for foreground and background regions and idea of edge prunning to delineate them as discussed in our previous work [1]. We use the labeled output and class probabilities map from a fully convolution neural network to improve the FG and BG seed points as well as edge estimates.…”