How effectively these systems can tell the difference between real and fake marks depends on how many factors are taken out and how they are changed. This is why the part of a detached mark check system that extracts parts is so important to how well the system works in general. The study shows a mixed technique for offline signature verification systems that pull-out features from pictures of signatures. The plan uses both the Convolutional Neural Network (CNN) and Histogram of Oriented Gradients (HOG) methods to find important features. The next step is to use a decision tree feature selection method. Two sets of data (UTSig and CEDAR) and three models (KNN, SVM, and LSTM) are operated to exam the mixed method again. The findings of the tests showed that it was very accurate at telling the variation among real and fake signatures, even for expert frauds.