H-PSO-LSTM: Hybrid LSTM Trained by PSO for Online Handwriter Identification

Abstract

The automatic writer’s recognition from his manuscript is a topical issue handling online writing. Recurrent neural networks (RNNs) are an effective means of solving such problem. More specifically, RNN networks with Long and Short Term Memory (LSTM) represent an ideal mean for writer’s recognition. Intuitively, LSTM networks are based on the gradient method for their learning processes. In addition, an LSTM node presents a complex data processing machine. Our hybrid approach combining LSTM and PSO (H-PSO-LSTM) presents the purpose of this paper and increases the performance of the network. Experiments were carried out on a Biometrics Ideal Test (BIT) bilingual database (Chinese and English). The BIT deals with a large number of writers (between 130 and 188). With H-PSO-LSTM, we were able to improve the learning performance accuracy to 91.9% instead of 81.2%.