基于连续非对称卷积结构的手写体数字识别

为了提高手写体数字识别的准确率,设计并提出了一种基于连续非对称卷积结构的手写体数字识别的深度学习算法.以连续非对称卷积结构为基础,结合极限学习机和MSRA初始化设计网络结构.在识别输入图像时,利用CUDA并行计算与Cudnn神经网络GPU加速库对手写体数字识别进行加速.在MNIST手写体数字数据库上进行实验,提出的网络结构识别准确率达到99.62%,单张图像识别速度为0.005 8 s.经实验结果对比表明,该网络结构在识别准确率和识别速度上得到有效提升.

Handwritten numeral recognition based on continuous asymmetric convolution structure

In order to improve the accuracy of handwritten numeral recognition, a deep learning algorithm for handwritten numeral recognition based on the continuous asymmetric convolution structure was designed and proposed. Based on the continuous asymmetric convolution structure and in combination with the extreme learning machine and MSRA initialization, the network structure was designed. With identifying the input image, the CUDA parallel computing and the Cudnn neural network GPU acceleration library were used to accelerate the handwritten numeral recognition. The experiments were performed on the MNIST handwritten digital database. The accuracy of network structure recognition is 99.62%, and the single image recognition speed is 0.005 8 s. The comparison of experimental results shows that both recognition accuracy and recognition speed for the present network structure has been effectively improved.