Algorithm 1 Sample Compression Algorithm

Input:  n images with size $s\times t$, the compression ratio $\mu$ and the iteration number b. Output:  n compressed images with size $\mu s\times t$. ($x_i$ stands for the $i_{th}$ input image, $img\_sampl{e}_i$ and $img\_construc{t}_i$ represent the $i_{th}$ output image and its reconstruction form, respectively, and $img\_sampl{e}_i^v$ and $img\_construc{t}_i^v$ represent the $v_{th}$ pixel value in the $i_{th}$ output image and its reconstruction form, respectively). 1: initialize the $g\_loss$ to infinity, the $kernel\_size$ to $\sqrt{\frac{1}{\mu }}\times \sqrt{\frac{1}{\mu }}$, the $stride$ to $\sqrt{\frac{1}{\mu }}$, $padding$ to 0 and $bias$ to $False$. 2: for $k=0$ to $b-1$ do 3:    for $i=0$ to $n-1$ do 4:      $img\_sampl{e}_i$$\leftarrow Conv2d\left(x_i\right)$ 5:      $img\_construc{t}_i$$\leftarrow ReconstructionModule(img\_sampl{e}_i)$ 6:      $g\_loss$$\leftarrow {\sum }_{v=1}^{s\times t}{\left(img\_construc{t}_i^v-x_i^v\right)}^2\times \frac{1}{s\times t}$ 7:    end for 8:    $Conv2d$$\leftarrow Optimizer(g\_loss)$ 9: end for 10:$img\_sampl{e}_i$$\leftarrow Conv2d\left(x_i\right)$ 11:return$img\_sampl{e}_i$