Integrating Structures Residual and C_Inception with Data Augmentation to Improve the Performance of LeNet-5

Aiming at the problems of insufficient generalization ability and low training efficiency of the current LeNet-5 model, this paper tries to improve the performance of the model from two aspects: model improvement and data enhancement. Firstly, based on LeNet-5, a new convolutional neural network C_LeNet-5 is formed by incorporating Residual structure and C_ Inception module. Secondly, MNIST is expanded to a new dataset C_ MNIST by data augmentation techniques such as brightness adjustment, gaussian noise injection, erasing, rotation, scaling and translation. Finally, the recognition performance of LeNet-5 and C_LeNet-5 on the C_MNIST dataset is compared, and the experimental results show that the C_LeNet-5 model performs better than the original model in handwritten digit recognition task. C_LeNet-5 model has stronger feature extraction ability, more stable training and faster convergence.