结合余弦相关性的卷积网络识别汉字的方法

Method of Combining Convolutional Neural Network with Cosine Similarity Algorithm to Recognize Chinese Characters

为了提高卷积神经网络(Convolution Neural Network,CNN)的识别率,增强卷积网络的特征提取能力,使其在模糊、光照不均等恶劣条件下能够有更好的识别效果,因此提出将余弦相关性加入神经卷积网络作为相似度度量的方法。较传统神经卷积网络相比较,有着更强的模式检测能力、更快的收敛速度以及更高的准确率的优点。在卷积神经网络的卷积层加入余弦相似性度量,最后通过对比传统神经卷积网络方法和余弦相关性神经卷积网络在脱机手写汉字的识别实验,在进行20次实验后,得出了在相同训练参数以及相同层数的卷积神经网络上,基于余弦相关性的神经卷积网络在手写汉字数据集上的准确率比传统的神经卷积网络的识别率平均提高了2.01%,并且有着更快的收敛速度。最后通过与现今流行的算法在MNIST数据集上的实验进行准确度、损失函数、时间复杂度的比较,得出结合余弦的卷积神经网络在准确度和损失函数上有一定的优势性,在时间复杂度上还需进一步提高。

In order to improve the recognition rate and enhance feature extraction capability of Convolutional Neural Network(CNN),a method for CNN with cosine similarity algorithm is proposed.It has better recognition results under harsh conditions such as blurring and uneven illumination.So a cosine similarity algorithm is added to the convolutional layer of the CNN.The method has the advantages of stronger pattern detection ability,faster convergence speed and higher accuracy than traditional CNN.By adding a cosine similarity algorithm to the convolutional layer of the CNN,the handwritten Chinese characters recognition experiments based on the traditional CNN method and cosine similarity CNN method respectively are completed,with the same training parameters and the same number of layers.After 20 experiments,the results show that the recognition rate of cosine similarity CNN is 2.01%higher and convergence speed is faster than the traditional CNN.Finally,the accuracy,loos function and time complexity are compared with the current popular algorithms on the MNIST dataset.It is concluded that the Convolutional Neural Network(CNN)combined with cosine similarity algorithm has certain advantages in accuracy and loss function.But further improvements are needed in terms of time complexity.