Recognition enhancement by linear tournament verification

The recent significant enhancement of OCR systems recognition rates has been driven mainly by combining different feature sets or by adopting a voting scheme using multiple independent algorithms. Voting is effective but computationally expensive. A general framework for economical enhancement of recognition rate that focuses on a critical reordering of a few top candidates is described. After the execution of the base OCR algorithm (a three-layer neural network), a linear tournament verification is executed using one-to-one small network verifiers to improve the ordering of the top candidates. Thirty-four one-to-one verifiers were developed for the uppercase English alphabet. Fourteen of these use special features; however, the rest use the same features as those in the base algorithm. On the NIST uppercase data set, the recognition rate for the new system is 95.8%, showing a 1.2% improvement over the system without verification. Although the improvement is modest, the costs in both efficiency and development effort are small