基于边缘检测的卷积核数量确定方法

针对卷积神经网络中卷积核数量多凭经验确定的问题，提出了一种统计图像边缘信息来确定卷积核数量的方法。首先，采用边缘检测算子对训练图像进行边缘检测，并依据卷积层的卷积核尺寸对边缘图像进行边缘块提取；然后，统计提取到的边缘块以获得边缘特征矩阵；最后，计算边缘特征矩阵各列的方差，将方差排序且归一化，选择方差较大部分边缘类型的个数作为卷积核数量。在Mnist和Chars74K数据集上的实验结果表明，本文方法能依数据集特点自适应地确定卷积核数量，构造的卷积神经网络模型大小适应于特定数据集，且能获得较高分类准确率。

Method for determining the number of convolution kernel via edge detection approach

Conventionally, the number of convolution kernel is determined by experience. This paper proposes a method for determining the number of convolution kernel by counting the image edge information. First, the training image edges are detected. Then, the edge block is extracted based on the size of the convolution kernel. Second, the edge blocks are counted to obtain the edge feature matrix. Finally, the column variance of the edge feature matrix is calculated, normalized and sorted. The number of convolution kernel is then determined. The experimental results on the Mnist and Chars74K datasets show that the proposed method changes the number of convolution kernel adaptively for different data sets. Also, a higher classification accuracy is achieved.