Improving the lenet with batch normalization and online hard example mining for digits recognition
Yiliang Xie, Hongyuan Jin, Eric C.C. Tsang · 2017
Nowadays, applications based on digits recognition and characters recognition have become much more reliable thanks to the rapid development of the DNN(deep neural network) architecture and constantly increasing the efficiency to the computing resources. A lot of methods have been proposed to improve the performance of DNNs, such as the ReLU (Rectified Linear Unit) which is a widely used alternative of the sigmoid activation function, Local Response Normalization and Batch Normalization. What's more, some methods, OHEM (Online Hard Sample Mining) for instance, aims to augment the capability of generalization of DNNs in certain scenarios. In this paper, we choose the LeNet architecture as our baseline model, which is known to work well on digit classification tasks. Starting from this baseline model, our goal is to investigate its drawbacks and enhance it by applying three different strategies including: adding activation layers right after each convolutional layer, applying batch normalization and, finally, the online hard example mining is adopted. By combining these methods together, we boost the performance of the baseline model from 98.106% to 98.720%.