Algorithm 1. The Parameter Updating Algorithm

Input: Training images $X$ and the corresponding labels $Y$, parameters ${\theta }^0$ of embedding model, parameters $W^0$ of the classification model, balance parameter λ, learning rate lr, training epochs L. 1: i = 0 2: While i < L do: 3: i = i + 1, $X_{\mathrm{Batch}},Y_{\mathrm{Batch}}=\mathrm{Sampling}\left(X,Y\right)$, 4: ${\mathit{Z}}_{\mathrm{Batch}}=F\left(X_{\mathrm{Batch}}\right)$,${\hat{Y}}_{\mathrm{Batch}}=C\left(Z_{\mathrm{Batch}}\right)$ //Perform feature extraction and classification. 5: ${\mathit{Z}}_{\mathrm{p}},{\mathit{Z}}_{\mathrm{n}}=\mathrm{Select}\left(Z_{\mathrm{Batch}},Y_{\mathrm{Batch}}\right)$ //Select positive and negative samples. 6: ${\mu }_{\mathrm{p}}=\mathrm{Average}\left({\mathit{Z}}_{\mathrm{p}}\right)$,${\mathit{Z}}_{\mathrm{p}},{\mathit{Z}}_{\mathrm{n}}=\mathrm{Hard\; sample\; mining}(Z_{\mathrm{Batch}},{\mathit{Z}}_{\mathrm{p}},{\mathit{Z}}_{\mathrm{n}}),$ 7: $L_{\mathrm{CE}}^i=L_{\mathrm{CE}}\left({\hat{Y}}_{\mathrm{Batch}},Y_{\mathrm{Batch}}\right)$, $L_{\mathrm{CovD}}^i=L_{\mathrm{CovD}}\left(Z_{\mathrm{p}},Z_{\mathrm{n}},{\mu }_{\mathrm{p}}\right)$, $L_{\mathrm{Total}}^i=L_{\mathrm{CE}}^i+\lambda L_{\mathrm{CovD}}^i$, 8: ${\theta }^{i+1}={\theta }^i-lr\cdot \frac{\partial L_{\mathrm{Total}}^i}{\partial {\theta }^i}$, $W^{i+1}=W^i-lr\cdot \frac{\partial L_{\mathrm{CE}}^i}{\partial W^i}$.//Update the model parameters. 9: end while Output: ${\theta }^L$, $W^L$.