In this work, a recognition approach applicable at the letter block (subword) level
for Arabic manuscripts is introduced. The approach starts with the binary images
of the letter block to build their input representation, which makes it highly objective and independent of the designer. Then, using two different manifold learning
techniques, the representations are reduced and learned. In order to decrease the
computational complexity, PCA is applied to the input representations before manifold learning is applied. Also, in order to increase the performance and quality of
the input representations, a gray stroke map (GSM) is considered in addition to the
binary images