Segmentation-Verification for Handwritten Digit Recognition

Anglický jazyk

Segmentation-Verification for Handwritten Digit Recognition

Na objednávku

44.37 €

bežná cena: 49.30 €

O knihe

Doctoral Thesis / Dissertation from the year 2016 in the subject Computer Science - Applied, National Higher School Of Computer Engineering, language: English, abstract: Automatic reading of digit fields from an image document has been proposed in several applications such as bank checks, postal code and forms. In this context, two main problems occur when attempting to design a handwritten digit string recognition system. The first problem is the link between adjacent digits, which can be naturally spaced, overlapped or/and connected. The second problem is the unknown length of the digit string, which is not carefully written by people in real-life situations.

In this thesis, SVM-based segmentation-verification system for segmenting two connected handwritten digits using the oriented sliding window is proposed. It employs a segmentation-verification system using conjointly the oriented sliding window and Support Vector Machine (SVM) classifiers. Experimental results showed that the proposed system is more appropriate for segmenting simple and multiple connections. Its main advantage lays in the use few rules for finding the optimal segmentation path. Hence, the proposed approach constitutes a tradeoff between the correct segmentation and the number of the segmentation cuts.

Thereafter, we propose a new design of a handwritten digit string recognition system based on the explicit approach for the unknown-length digit strings. Three methods are combined according the link of adjacent digits, which are the histogram of the vertical projection dedicated for spaced digits, the contour analysis dedicated for overlapped digits and the Radon transform performed on the sliding window dedicated for connected digits. A recognition and verification module based on Support Vector Machine (SVM) classifiers allows analyzing and deciding the rejection or acceptance each segmented digit image. Experimental results conducted on the benchmark dataset show that the proposed system is effective for segmenting handwritten digit strings without prior knowledge of their length comparatively to the state-of-art.