# Image reconstruction using Laplace pyramid

Given a sub-sampled Image $P_n$ belonging to layer $n$, a Gaussian kernel $G$ and a Laplace Layer $L_{n-1}$ of layer $n-1$, my efforts to create $P_{n-1}$ of previous layer $n-1$ have failed.

This was what I attempted.

1. Created a staging image $S_{n-1}$ by padding alternate row and column of $P_n$ with $0$.
2. Convolved $S_{n-1}$ with $G$ resulting in say $B_{n-1}$. This was done with the intention of interpolating and replacing the $0$s with approximated values.
3. Added $L_{n-1}$ (Missing Information) to $B_{n-1}$ hoping that the result would resemble the original $P_{n-1}$.

Two Gaussian kernels ($3\times 3$ and $5\times 5$ with $\sigma=1$) were used in my application.

The resulting image is hopelessly dark.

• Are the steps 1-3 correct?
• Should I use an alternative interpolation technique such as cubic instead of convolution in Step-2 before adding the Laplace layer?
• Do you need the "Recipe" of Laplacian Image Pyramid? – Royi Dec 7 '16 at 18:02
• Make sure that your image pixel types can accommodate a signed image, as L is typically made up of both positive and negative values. – mikeTronix Dec 7 '16 at 19:44
• Is B(n-1) always darker than P(n)? That is, do they have very different average values? If so, your interpolation kernels are not chosen properly. – mikeTronix Dec 7 '16 at 19:50
• @Royi, Not sure if I understood your comment. But if it meant "Do you require precise steps", that would be very welcome for a student like me. Best wishes. – Raj Dec 14 '16 at 5:41