Algorithm 1 Random Local Rotation

1:Input: The training dataset $\mathcal{D}$  2:Output: The augmented training dataset $\tilde{\mathcal{D}}$  3:Create a subset say ${\mathcal{D}}^{\prime }\subseteq \mathcal{D}$ by randomly selecting N images from $\mathcal{D}$  4:Create an empty set say ${\mathcal{D}}^{*}$  5:for$\forall \mathit{I}\in {\mathcal{D}}^{\prime }$do  6:    Randomly generate centre $(x,y)$ and radius r within image I  7:    Form a circular area $C_{x,y,r}$ within image I with centre $(x,y)$ and radius r  8:    Randomly generate angle $\theta \in [0,2\pi ]$  9:    Form image $\tilde{\mathit{I}}$ by rotating the area within $C_{x,y,r}$ in image I with angle $\theta$ 10:   Add $\tilde{\mathit{I}}$ into ${\mathcal{D}}^{*}$ 11:end for 12:Way 1: $\tilde{\mathcal{D}}\leftarrow {\mathcal{D}}^{*}\cup \mathcal{D}\setminus {\mathcal{D}}^{\prime }$ 13:Way 2: $\tilde{\mathcal{D}}\leftarrow {\mathcal{D}}^{*}\cup \mathcal{D}$