Off-line writer identification using multi-scale local binary patterns and SR-KDA

Writer identification is becoming an increasingly important research topic especially in forensic and biometric applications. This paper presents a novel method for performing offline write identification by using multi-scale local binary patterns histogram (MLBPH) features. The proposed feature (MLBPH) when combined with edge-hinge based feature achieves a top 1 recognition rate of 92% on the benchmark IAM English handwriting dataset, outperforming current state of the art features. Further, kernel discriminant analysis using spectral regression (SR-KDA) is introduced as dimensionality reduction technique to avoid the overfitting problem associated with using multi-scale data.