Algorithm 1 FW-RFE

Input: The set of output features of the pooling layer and sample labels, $${\left\{(P_i,y_i)\right\}}_{(i=1)}^{Num\_l},P_i\in {\left\{p_1,p_2,\cdots ,p_j\right\}}_{j=1}^{Num\_f},y_i\in \left\{+1,0\right\}$$    and the set of convolutional layer feature weights, $${\left\{\left[{w_1}^1,{w_2}^1,\cdots {w_{10}}^1\right],\left[{w_1}^2,{w_2}^2,\cdots {w_{10}}^2\right],\cdots ,\left[{w_1}^j,{w_2}^j,\cdots {w_{10}}^j\right]\right\}}_{(j=1)}^{Num\_f}$$ Output: Feature Sorted Set R; 1: Initial feature set S $$S={\left\{s_i\right\}}_i^{num\_l},s_i=\left\{p_1^i,{p_2}^i,\cdots ,{p_j}^i\right\}$$ and accuracy set $V=⌀$; 2:while$(S\ne \varphi )$do 3:    Calculate the ranking criterion score:$c_i={\displaystyle \sum _{n=1}^{10}}{\left(w_n^j\right)}^2$; 4:     Find the feature with the smallest ranking score: $$h=minC$$ 5:    Extract features $s_i=s_i-p_h^i$, updata S; 6:    Take S as the input of the fully connected layer, use sigmoid regression function to achieve binary classification: $sigmoid\left(S_i\right)=\frac{1}{1+e^{-S_i}}$ 7:    Update accuracy set: $V=\left[acc,V\right]$ 8:end while 9:When the accuracy rate in the accuracy set V is the highest, the feature set in S is selected as the output result