Multiactivation Pooling Method in Convolutional Neural Networks for Image Recognition

Table 7

The baseline DenseNet-40 architecture ‎[15] and our DenseNet-40_MAP architectures. Two models are both designed for CIFAR-10 datasets.

Layers	Densenet-40	Densenet-40_MAP
Convolution	conv3 stride 1 padding 1
Dense Block(1)	[BN-ReLU-conv3] 12
Transition Layer(1)	[BN-ReLU]
	conv1	[conv3-ReLU] 3
	avgpool 22	MAP 44
Dense Block(2)	[BN-ReLU-conv3] 12
Transition Layer(2)	[BN-ReLU]
	conv1	[conv3-ReLU] 3
	avgpool 22	MAP 44
Dense Block(3)	[BN-ReLU-conv3] 12
Classification Layer	avgpool 88	avgpool 22
	10D fully-connected, softmax