Text binarisation process classifies individual pixels as text or background in the textual images. Binarization is necessary to bridge the gap between localization and recognition by OCR. This paper presents Sliding window method to binarise text from textual images with textured background. Suitable preprocessing techniques are applied first to increase the contrast of the image and blur the background noises due to textured background. Then Edges are detected by iterative thresholding. Subsequently formed edge boxes are analyzed to remove unwanted edges due to complex background and binarised by sliding window approach based character size uniformity check algorithm. The proposed method has been applied on localized region from heterogeneous textual images and compared with Otsu, Niblack methods and shown encouraging performance of the proposed method.
Deep Dive into Sliding window approach based Text Binarisation from Complex Textual images.
Text binarisation process classifies individual pixels as text or background in the textual images. Binarization is necessary to bridge the gap between localization and recognition by OCR. This paper presents Sliding window method to binarise text from textual images with textured background. Suitable preprocessing techniques are applied first to increase the contrast of the image and blur the background noises due to textured background. Then Edges are detected by iterative thresholding. Subsequently formed edge boxes are analyzed to remove unwanted edges due to complex background and binarised by sliding window approach based character size uniformity check algorithm. The proposed method has been applied on localized region from heterogeneous textual images and compared with Otsu, Niblack methods and shown encouraging performance of the proposed method.
Understanding texts from natural scenes/scene text image such as, commercial signboards, traffic signs, and advertising billboards is very useful in many purposes such as assistant system for impaired persons, drawing attention of a driver to traffic signs, text translation system for foreigners, potential applications like license plate recognition, digital note taking, document archiving and wearable computing. Binarization problem concerns classifying individual pixels as text or background. Binarization is necessary to bridge the gap between localization and recognition by OCR. The output of this step is a binary image where black text characters appear on a white background. Current techniques are categorized into two groups: global binarization and local binarization or adaptive binarization.
In global binarization methods [12], global thresholds are used for all pixels in image and are not suitable for complex and degraded document images. Global methods are fast and robust for small text. In the other hand, local binarization methods change the threshold adaptively over the image according to properties of local regions. Local binarization methods are proposed to overcome binarization drawbacks in global ones. Local binarization methods can be improved by calculating local thresholds within separate windows or areas [11] [5]. In most of these methods, the size and shape of the window are predefined parameters. Poor binarization results are obtained when a window’s boundaries cross characters and may give rise to broken characters and voids, which may cause undesirable artifacts in the binary image.
And other challenging issue related to binarising text from textual images is the presence of complex/textured background. Here sliding window approach based binarisation method is proposed to binarise the text from color documents with textured /complex background. The paper is organized as follows. Section 2 deals with prior and related works, Section 3 illustrate our method with various modules in Section 4-6. In Section 7, experimental results are reported and conclusions and future works are summarized in Section 8.
Various text binarization techniques have been found in literature and are discussed here. Dual binarization method is proposed in [1] which can easily segment texts with different two color polarities from backgrounds in the key caption area. [10] Proposed a binarisation method to remove the background pixels inside the characters also. The final binary image is gotten by fusing the three binary images such as the locally adaptive seed-fill method, the locally adaptive thresholding method and the stroke-model-based method. A learning-based binarization method is proposed in [3] for same-type documents. In this paper, the stroke width is used to evaluate the binarization .This approach can be used for only same type of documents. In [4] technique based on a Markov Random Field (MRF) model of the document is proposed. The model parameters (clique potentials) are learned from training data and the binary image is estimated in a Bayesian framework.
In [5] Segmentation of text in the detected text region is performed with all color components into two distinctive colors to discriminate between text and other non-text region with fuzzy c-mean (FCM) clustering to depict the color distribution. Adaptive local thresholding based on a verification-based multi threshold probing scheme is proposed in [6]. This approach is regarded as knowledge-guided adaptive thresholding. [7] proposed a text segmentation method based on spectral clustering and the histogram of intensity is used for the object of grouping. This algorithm uses the normalized graph cut measure as the thresholding principle to distinguish an object from the background.
However these methods mainly work on the images of text with nearly uniform background. Some works deal with complex background [8] [9] [13] [9] and on each word [8]. But they assume the text color to be uniform and not suitable for color documents with multicolored text. [2] deals with multicolored text documents regardless of the polarity of foreground-background shades with edge-box analysis. However, if the background is textured, the edge components may not be detected correctly due to edges from the background and edge-box filtering strategy fails.
Therefore, it is proposed to address the above issues by proposing the Sliding window based character size uniformity check algorithm to minimize the complexity of background.
Here, an approach is proposed in which iterative thresholding is used to detect edges instead of fixed global threshold which will usually work well for images with uniform background, but not for textured background. Here uniformity of character sizes is analyzed to remove false edges due to textured background. Proposed binarization technique consists of the following processes: Preprocessing, Iterative thresholding for edge detection,
…(Full text truncated)…
This content is AI-processed based on ArXiv data.
