Document Image Binarization Based on NFCM

In general, binarization methods of document image can be categorized into two classes [5]: local methods and global methods. The global threshold methods segment an entire image with a single threshold using the gray level histogram of image, while the local methods partition the given image into a number of sub-images (local windows) and select a threshold for each of sub-images. The most common methods of calculating the global threshold are Otsu’s algorithm [6] and the histogram shapebased threshold algorithms [6]. The global threshold method has a very good effect for the document image which has two obvious peaks in gray level histogram or is clearly between the foreground and the background. However, for the document image when the gray values between the foreground and the background become fuzzy or whose brightness is uneven, it can not obtain a good result. The most common methods of calculating local threshold are Bernsen algorithm [6] and Niblack algorithm [6]. The local threshold method is fit for dealing with the situation of the edge between character and the background being relatively vague and uneven illumination. In this case, the local threshold approach can extract character strokes from the background effectively. However, the local threshold methods also have some shortcomings. Firstly, when the whole local window is filled up with the background, the local binarization methods usually bring the ghost phenomenon (ghost is a pseudo stroke which is derived by binarization algorithm in the background region) [7]. In addition, the normal local threshold methods will also bring fracture phenomenon for some camera images (the strokes in the characters appears fracture phenomenon) [8]. Furthermore, some of the local threshold methods are relatively slower. Recently, some new binarization methods were proposed. [9] proposed a method which is based on mathematical morphology operation to enhance the blurred stroke information. [10] proposed a method which is a combination of Otsu and Bernsen. [11] proposed binarizing document image with K-means clustering algorithm. However, the K-means algorithm is a hard clustering algorithm. All these papers are not based on the combination of Niblack and Fuzzy C-Means algorithm (FCM), which is a soft clustering algorithm and has higher ability to process uncertain information [12], e.g. the fuzzy document image. Motivated by the aforementioned reason, we propose a new binarization algorithm for camera-based document image. The algorithm includes two thresholds: the first local threshold is calculated by Niblack algorithm; the second local threshold is calculated by FCM. We call this algorithm NFCM algorithm. This algorithm can not only preserve the character stroke details, but also eliminate ghost effectively.