Comparative Analysis of Recent Architecture of Convolutional Neural Network

Table 2

Comparison of CNN architecture.

S.NO	Architecture	Year	Major contribution	Parameters	Depth	Reference
1	LeNet	1998	Initial CNN architecture	0.060 M	5	[15]
2	AlexNet	2012	Deeper than LeNet and uses relu and overlapping pooling	60M	8	[20]
3	ZfNet	2014	Visualizing intermediate layer	60 M	18	[21]
4	VGG	2014	Uses small kernel size	4M	19	[22]
5	Google Net	2015	Introduced block concept, split transform, and merge idea	23.6 M	22	[23]
6	Inception V3	2015	Bottleneck issue is sorted and small filter size is added	23.6 M	159	[24]
7	Highway networks	2015	Introduced multipath concept	2.3 M	19	[25]
8	Inception V4	2016	Split transform and merge idea	35M	70	[26]
9	ResNet	2016	Residual learning	25.6	152	[27]
10	Deluge Net	2016	Allows cross layer information	20.2 M	146	[28]
11	Xception	2017	Depth-wise convolution followed by a point-wise convolution	8.6 M	452	[29]
12	ResNeXt	2017	Cardinally homogenous topology grouped convolution	27.5 M	152	[30]
13	Dense Net	2017	Cross-layered information flow	25.6 M	190	[31]
14	Channel boosted CNN	2018	Boosting of the original channel with additional artificial channels	—	—	[32]
15	Convolution block attention module	2018	Feature-map attention	—	—	[33]