Incremental MQDF Learning for Writer Adaptive Handwriting Recognition

Abstract

Writer adaptation has been proved to be an effective approach to improve the recognition performance of the writer-independent recognizer for a particular writer. In this paper, we propose a writer adaptive handwriting recognition approach by incremental learning the Modified Quadratic Discriminant Function (MQDF) classifier. We derived the solution of Incremental MQDF (IMQDF) and then present a Discriminative IMQDF (DIMQDF) by deriving the solution of IMQDF in the updated discriminative feature space. Based on IMQDF or DIMQDF, the writer adaptation is finally performed by updating the MQDF recognizer adaptively. The experimental results for recognizing handwriting Chinese characters indicate that the proposed IMQDF and DIQMDF approaches can reduce as much as 52.71% and 45.38% error rate respectively on the writer-dependent dataset while only have less than 0.18% accuracy loss on the writer-independent dataset. In other words, the proposed IMQDF and DIMQDF based writer adaptation approaches can significantly increase the recognition accuracy on writer-dependent dataset while only have limited negative influence for general writer.

Writer Adaptation with Style Transfer Mapping

“…is the Mahalanobis distance of class i, and d 2 ðx; iÞ is equal to the first two terms of (19). The hyperparameter !…”

“…where Gðx; iÞ can be either (15) or (19), which can be viewed as a distance metric between x and class i. The top-1 output is then defined as the distance between x and its predicted classŷ, while the top-2 output is the distance from the most competing class:…”

“…For describing these adaptation procedures, we use Gðx; kÞ to represent either G LVQ ðx; kÞ in (15) or G MQDF ðx; kÞ in (19), which is the distance metric between a pattern x and class k. The target point function is denoted by T ðx; yÞ, which can be either T LVQ ðx; yÞ in (17) or T MQDF ðx; yÞ in (22). The confidence function is F ðxÞ in (28).…”

“…The number k of principal axes for MQDF in (19) acts as a balance between the memory requirement and the classification accuracy. We compared the performance of adaptation by varying k ¼ 10; 30; 50 (denoted by MQDF(10), MQDF(30), MQDF(50)).…”

“…For Bayesian classifier adaptation, Takebe et al proposed using the difference distribution to adapt the quadratic discriminant function (QDF) to a new font [18]. Ding and Jin proposed the adaptation of the modified quadratic discriminant function by incremental learning [19].…”