Pooling layers are spatial down-sampling layers used in convolutional neural networks (CNN) to gradually downscale the feature map, increase the receptive field size and reduce the number of the parameters in the model. The use of pooling layers leads to less computing complexity and memory consumption reduction but also introduces invariance to certain filter distortions which may induce subtle detail loss. This behaviour is undesired for some fine-grained recognition tasks such as facial expression recognition (FER) which highly relies on specific regional distortion detection. In this paper, we introduce a more filter distortion aware pooling layer based on kernel functions. The proposed pooling reduces the feature map dimensions while keeping track of the majority of the information fed to the next layer instead of ignoring part of them. The experiments on RAF, FER2013 and ExpW databases demonstrate the benefits of such layer and show that our model achieves competitive results with respect to the state-of-the-art approaches.

池化层是卷积神经网络（CNN）中使用的空间下采样层，用于逐步缩小特征图的比例，增加接收场的大小并减少模型中参数的数量。池化层的使用导致较少的计算复杂性和内存消耗的减少，但也对某些滤波器失真引入不变性，这可能引起细微的细节损失。对于某些细粒度的识别任务（例如面部表情识别（FER）），这种行为是不希望的，而面部表情识别（FER）高度依赖于特定的区域失真检测。在本文中，我们介绍了一个基于内核函数的更具过滤器失真意识的池化层。拟议的合并减少了特征图的尺寸，同时跟踪了馈送到下一层的大多数信息，而不是忽略了其中的一部分。