Improvement Accuracy Identification and Learning Speed of Offline Signatures Based on SqueezeNet with ADAM Backpropagation
Yosepin Petra Purbanugraha, Adian Fatchur Rochim, Iwan Setiawan · 2022
Signatures play an essential role in human life because they are used as an authentication approach in organizations such as banks, businesses, and legal authorities. In offline signature identification, dynamic information such as velocity of the stroke, hand vibration, signature image position and pen pressure from the signature writing process is lost, so offline analysis has less information and as a result the identification becomes more complicated. Currently, from various websites, there are many electronic documents that contain offline signatures. This leads to the problem of offline signature identification. Offline signature identification can be viewed as part of the Fine-Grained Visual Classification (FGVC) problem. FGVC research is a challenge because there are problems with fine-grained datasets and it can quickly become a big problem due to factors such as angle of view, tilt, and the location of objects in the image. In recent years, Convolutional Neural Networks (CNN) are a powerful and very successful approach for identification purposes. In this research, a CNN with small networks architecture namely SqueezeNet is optimized using Adaptive Moment Estimation (ADAM) to achieves high accuracy and learning speed, it is trained using four public signature datasets, namely BHSig260-Bengali, BHSig260-Hindi, CEDAR, UTSig and then it is tested to identify the original signature samples taken from each dataset. The optimized SqueezeNet achieves better results compared with other deep learning or machine learning algorithms. The optimized SqueezeNet takes under 3 minutes to train networks for each dataset and its identification average accuracy reaches above 99,79% for each dataset.