Thai handwritten character recognition using heuristic rules hybrid with neural network

Abstract

This research enhanced two major processes of the previous work of the off-line Thai handwritten character recognition using hybrid techniques of heuristic rules and neural network system. The proposed functions are mainly in 1) Feature extraction enhancement to improve the feature conflict resolution rule and the specialized neural network-based zigzag feature extraction. These functions are used to refine the conflict features and zigzag patterns; 2) Neural network-based recognition. Specifically, a neural network technique improves the capabilities of the recognition process to handle various styles of writing. The result showed that the additional feature conflict resolution rule could achieve the feature extraction rate of 87.85% (increased 2.13%), the feature extraction rate of the specialized neural network-based zigzag extraction could achieve 90.48% (increased 47.9%) and the recognition rate of the neural network-based recognition which combine both of the two proposed feature extraction functions could achieve 92.78% (increased 9.77%).

Recognition of handwritten characters using local gradient feature descriptors

“…In one related paper (Nopsuwanchai et al, 2006), their method obtained 92.03% on the ThaiCAM database. In another paper (Mitrpanont and Imprasert, 2011), their method obtained 92.78% on the Thai handwritten character dataset. Our best method obtains 94.34% on the test set and 98.93% with cross validation on the THI-C68 dataset and 97.87% on the test set and 99.07% with cross validation on the THI-D10 dataset.…”

“…They obtained 92.03% accuracy on the ThaiCAM database. Some hybrid techniques of heuristic rules and neural networks are employed in Mitrpanont and Imprasert (2011). The performance obtained from this approach on the Thai handwritten character dataset was 92.78%.…”

Zig-zag diagonal and ANN for English character recognition

“…The zig-zag scan is one of the most important for complex patterns, and it achieved good accuracy in Iris recognition [23]. Mitrpanont and Imprasert [24] believed that zig-zag method is a useful way to extract the features of those characters who have similar features and authors applied it to extract the features of handwritten Thai characters and increased accuracy was 2.13% from the other existing system. So, according to this analysis, it could be suggested that the zig-zag diagonal method along with zoning and binarization will perform well to extract the features of English character and digits and will help to increase the accuracy.…”

Thai scene text recognition

“…Finally, the recognized characters were grouped into word level and corrected by the predefined dictionary. (Mitrpanont & Imprasert, 2011) presented a novel zig-zag feature for Thai handwritten recognition. First, the zig-zag feature was extracted from a 64x64 pixel segmented character image in a vertical direction, acquiring 64x1 vector fed into zigzag and non-zigzag MLP classifier.…”