An Investigation of Feature Selection and Transfer Learning for Writer-Independent Offline Handwritten Signature Verification

SigNet is a state of the art model for feature representation used for\nhandwritten signature verification (HSV). This representation is based on a\nDeep Convolutional Neural Network (DCNN) and contains 2048 dimensions. When\ntransposed to a dissimilarity space generated by the dichotomy transformation\n(DT), related to the writer-independent (WI) approach, these features may\ninclude redundant information. This paper investigates the presence of\noverfitting when using Binary Particle Swarm Optimization (BPSO) to perform the\nfeature selection in a wrapper mode. We proposed a method based on a global\nvalidation strategy with an external archive to control overfitting during the\nsearch for the most discriminant representation. Moreover, an investigation is\nalso carried out to evaluate the use of the selected features in a transfer\nlearning context. The analysis is carried out on a writer-independent approach\non the CEDAR, MCYT and GPDS datasets. The experimental results showed the\npresence of overfitting when no validation is used during the optimization\nprocess and the improvement when the global validation strategy with an\nexternal archive is used. Also, the space generated after feature selection can\nbe used in a transfer learning context.\n