Fig. 2.

Proposed structural units. Here, $\mathbf{h}$, $\mathbf{w}$, and $\mathbf{c}$ denote the height, width and no. of channels of the feature map, respectively, while ‘$\mathbf{k}$’ stands for kernel size, ‘$\mathbf{s}$’ for strides and ‘$\mathbf{f}$’ for number of filters in the convolution. In depthwise convolution, dilation rate will be varied from $\mathbf{1}$ to ‘$\mathbf{m}$’.