基于Sobel算子的池化算法设计

池化算法是卷积神经网络中用于特征降维、参数压缩、扩大感受野的重要一层。针对现有的池化方法没有充分考虑到池化前特征图的整体内容及风格特征分布问题,提出了一种通过Sobel算子对卷积后的特征图计算每个特征点的梯度值,并根据梯度值分布确定每个池化窗口取最大值、均值或者最小值的池化算法。该算法充分考虑了特征图池化前后的整体内容及风格特征分布,保持了特征图的整体不变性。实验表明,该池化算法在VGG、ResNet等经典网络架构上取得了优异性能,具有普适性,可用来替代常用的最大池化、平均池化。

Design of Pooling Algorithm Based on Sobel Operator

The pooling algorithm is an important layer in the convolutional neural network for feature dimension reduction, parameter compression, and expansion of the receptive field. Aiming at the problem that the existing pooling methods do not fully consider the overall content and style feature distribution of the feature map before pooling, A method is proposed to calculate the gradient value of each feature point on the convolved feature map by the Sobel operator, and according to the gradient value distribution, determine the pooling algorithm for each pooling window to take the maximum value, the average value or the minimum value. The algorithm fully considers the overall content and style feature distribution before and after the feature map pooling, and maintains the overall invariance of the feature map. Experiments show that the pooling algorithm achieves excellent performance on classic network architectures such as VGG and ResNet, and is universal, can be used to replace the commonly used max pooling and average pooling.