Algorithm 1
1.	Input all the image patches in the sample set X with the pre-whitening steps and initialize the basis function matrix A
2.	Calculate the mean vector m from the input image patches, $\mathit{\text{m}}=\frac{1}{n}\sum _{i=1}^n{\mathit{\text{x}}}_i$
	Construct the d × n mean matrix for the sample set X, M =(m, m,…,m)
	Apart from the mean matrix and get the centered sample matrix X̅, X̅ = X−M
3.	Get a d × d covariance matrix, C = X̅X̅T = UΛUT with the diagonal matrix of the eigenvalues Λ = diag[λ1,…λi,…,λd] and the eigenmatrix U = [u1,…,ui,…,uk].
	Adjust Λ in a descending order and get the corresponding eigenvector ui by the arrangement of the eigenvalue λi
	Apply the PCA projection and select the first qth eigenvectors to form the whitening matrix V in a q × d size, $\mathbf{\text{V}}={\mathbf{\Lambda }}^{-\frac{1}{2}}{\mathbf{\text{U}}}^{\text{T}}$, where Λ turns to be a q × q matrix and U a q × d matrix.
4.	Whiten the centered sample matrix X̅ with Z = VX̅ in a q × n size.
5.	Take the Fast-ICA algorithm and get the orthogonal matrix B = (b1,…bi,…,bq) in a q × qsize, A = V−1B, where the basis function A is a d × q matrix and the i th element bi in B is a q-dimensional vector. Compute the ICA feature coefficient matrix S = BTZ in a q × n size.
6.	Reconstruct the approximation of the original image sample set X̃ with X̃ = AS.
7.	Whiten again by Step 3 and get the d × n whitened matrix S by S= VTS.
8.	Discard the signs of the output S in this layer and take the nonlinear activation function g to form the input sample set X′ for the next layer, X′=g(|S|).
9.	Repeat the above process to achieve the recursive multiple layer ICA architecture