Improved Handwritten Numeral Recognition on MNIST Dataset with YOLO and LSTM

With the aging of population and the advance of technology, handwritten numeral recognition system is sophisticated and widely used. However, due to the presence of different writing surfaces, postures and other factors, the performance of handwritten numeral recognition is limited. In this paper, we propose a new supervised recurrent neural network, which combines time and space for target location prediction on handwritten datasets. Our method is based on the YOLO framework, and combines a long and short term memory (LSTM) mechanism. Moreover, our method not only locates handwritten images, but also improves the classification accuracy. Extensive comparison with the state-of-the-art methods demonstrates that our method achieves both accuracy and robustness on handwritten datasets. Meanwhile, our method is effective with low computational cost.