TY  - THES
T1  - Offline handwritten word recognition system using PRG and MLP-HMM hybrid classifier
A1  - Cajote, Rhandley D.
LA  - English
YR  - 2004
UL  - https://tuklas.up.edu.ph/Record/UP-99796217607612508
AB  - Offline handwritten word recognition is one of the most difficult problems in the area of pattern recognition. The large variability and ambiguity in the two dimensional pattern of handwritten words makes it difficult to model the handwriting process. An optimal feature set that describes handwriting is still a subject of research. The lack of a standard database to test and compare algorithms is still a subject of research. The lack of a standard database to test and compare algorithms adds to the difficulty of the problem. Despite all these, significant progress has been made in offline handwriting recognition in the areas of mail address and bank check recognition. The techniques in speech recognition have been used effectively in handwriting recognition. These include modeling the process of handwriting as a sequence of left-right signals analogous to the temporal sequence in speech and using a hybrid MLP-HMM recognition system. In this thesis an offline cursive handwritten word recognition system was developed. The recognition system uses local and global features with an MLP-HMM hybrid classifier. Using the local features alone and an MLP-HMM hybrid recognizer the highest recognition rate achieved is 70% using a 20-word vocabulary. The global features m based on the PRG of the word contour, combined with the local geometric features have been found to be successful in improving the performance of a recognition system. The combined PRG-HMM-MLP recognizer is able to achieve a recognition rate of 72% for a 20-word vocabulary tested using the demo version of the IAM database. Higher recognition rates can still be achieved by the system if more training data is available.
CN  - LG 995 2004 E64 C35
KW  - Writing : dentification.
KW  - Pattern recognition systems.
KW  - Pattern perception.
ER  -