卷积核初始设置对卷积神经网络表格识别的优化效果研究

Study on the Optimization Effect of Initial Setting of Convolution Kernel on Table Recognition of Convolutional Neural Network

由于卷积神经网络强大的特征学习与分类能力在图像识别与分类、目标检测、语义分割等领域得到了广泛地应用,提高卷积神经网络进行深度学习的效率和精度就显得更为重要。在卷积神经网络中,卷积核初始设计对深度学习的迭代效率等有着重要影响。本文以图像表格的识别与分类作为研究对象,提出了以图像表格的若干局部元素为基点,及其像素分布特征与初始化卷积核内参数分布相似的原则,对卷积核的初始化进行自定义设定卷积核参数,在此基础上进行图像表格的卷积神经网络深度学习,并与传统的Normal、Xavier等初始化方法进行了比较实验。实验结果表明,在神经网络学习过程中,本文的参数初始化方法在训练初期对表格识别的分类精度明显较高,总体分类准确率也明显较高。

Since the powerful feature learning and classification capabilities of convolutional neural network have been widely used in image recognition and classification,target detection,semantic segmentation and other fields,it is more important to improve the efficiency and accuracy of convolutional neural network for deep learning.In a convolutional neural network,the initial design of the convolution kernel has an important influence on the iteration efficiency of deep learning.Based on the identification and classification of image form as the research object,the image is put forward to form a number of local elements as the basis,and the pixel distribution feature similar to initialize the convolution kernel parameter distribution in principle,the initialization of convolution kernels to set custom convolution kernel parameters,on the basis of depth image form of convolution neural network learning,and compared with the traditional Normal,Xavier initialization method on the basis of the comparison experiment.The experimental results show that in the learning process of neural network,the parameter initialization method presented in this paper has a significantly higher classification accuracy in table recognition at the beginning of training,and the overall classification accuracy is also significantly higher.