Table 11.

Ablative analysis of components on D1 using average classification accuracy and computation complexity, where l, c, s, and k represents the corresponding layer of deep learning, number of input channels, and spatial size of output feature map, respectively. Bold emphasis indicate the best results

Component	Accuracy (%)	Computational complexity
VGG-16 with attention module	45.20	$\mathcal{O}\left(2.c_{l-1}\mathrm{.}{k}_l^2\right)+\mathcal{O}\left(c_{l-1}\mathrm{.}{s}_l^2\mathrm{.}{k}_l^2\right)$
VGG-16	75.25	$\mathcal{O}\left(c_{l-1}\mathrm{.}{s}_l^2\mathrm{.}{c}_l\mathrm{.}{k}_l^2\right)$
VGG-16 with both attention and convolution module (proposed model)	79.58	$\mathcal{O}\left(2.c_{l-1}\mathrm{.}{k}_l^2\right)+\mathcal{O}\left(c_{l-1}\mathrm{.}{s}_l^2\mathrm{.}{k}_l^2\right)$+ $\mathcal{O}\left(c_{l-1}\mathrm{.}{s}_l^2\mathrm{.}{c}_l\mathrm{.}{k}_l^2\right)$