Offline Signature verification using CNN and SVM classifier
B. H. Shekar, Wincy Abraham, Bharathi Pilar · 2022
An efficient signature verification method using both CNN and SVM is proposed. Cedar dataset consisting of a total of 2640 signature images of 55 individuals is used here as the dataset. Since it is a writer-dependent method, the genuine and forged signature images of each individual are used separately for training and testing. CNN (Convolutional Neural Network) is used here only for feature extraction and SVM is used for classification. The input images are supplied to the CNN input layer and the feature extracted, which is the downsampled convolution of the 9X9 kernel initialized to 1, sliding across the image. The extracted features are supplied to the Support Vector Machine (SVM) classifier in the 80 : 20 ratio and the classification result is found to have the best accuracy of 93.63%. The same features are used with CNN itself for classification to compare the performance. It gives a lower accuracy than expected because of the overfitting problem of the deep neural network. Thus use of CNN for feature extraction and SVM for classification yields good results in offline signature verification.