Table 9.

Runtime analysis of the attention-based CNN models considered for comparison in Table 7. To maximum possible extent, in most experiments the backbone was uniformly chosen to be ResNet50 in order to enable comparison of different attention approaches on top of the same CNN.

S. No.	Method	Backbone	Inference time (milliseconds/image)	Number of Parameters (in millions)	Giga FLOPs
1	FocusNet [34]	SE-Net50	12.38	26.82	2.74
2	Dual Attention Network [35]	ResNet50	35.45	49.51	14.27
3	Asymmetric Non-local networks [36]	ResNet50	52.78	44.04	12.57
4	Multi-scale self-guided attention [37]	ResNet50	60.73	38.78	10.19
5	Criss Cross Attention [38]	ResNet50	25.14	28.18	6.32
6	Semi Inf Net [8]	Res2Net	44.23	33.12	7.36
7	Proposed CNN	Inception-ResNet-V2 based MKE module	38.24	30.51	13.78