Multiple Layers Global Average Pooling Fusion

We propose for deep convolutional neural network (CNN) a simple but effective feature fusion technique called multiple layers global average pooling fusion (MLGAPF). It adds a branch at each CNN layer or module which uses global average pooling to extract global features, and these features are then fused for classification. Empirical experiments show that this technique can effectively improve the accuracy of ResNet, GoogleNet, SqueezeNet, MobileNetv2 and others. Onaverage, MLGAPF brought additional performance enhancement of 2.62% and 2.49% on CIFAR100 and Tiny-ImageNet respectively.