zxyctn / least-squares-correlation

MATLAB implementation of Least Squares Correlation/Matching (LSM) with grey value differences and gradients along the axes. Implemented as the final project for Photogrammetric Computer Vision course at Bauhaus-Universität Weimar.

Least-Squares-Correlation

MATLAB implementation of Least Squares Correlation/Matching (LSM) with grey value differences and gradients along the axes. Implemented as the final project for Photogrammetric Computer Vision course at Bauhaus-Universität Weimar.

Description

Algorithm takes a greyscale 150x150 image as an input and with an arbitrarily chosen affine deformation matrix produces a distorted image to be iteratively re-transformed back to its initial state with LSM. After fine-tuning the parameters, a target image and its 3 deformed versions are tested against the system for evaluation. Images can be found in the img folder. Number of iterations are manually adjusted for the best result. To avoid unnecessary void that may be caused by the deformations, images are cropped in center by 100x100 size.

Arbitrarily chosen deformation matrices

In the main.m file, there are 4 matrices to be applied and manually chosen. The matrices are stored in the H_0 variable. To test with the source image and application of transformation matrices, refer to the comments in the file.

Evaluation

Referring to the comments and instructions in the main.m file, you'll be able to take the target image and the 3 distorted images as inputs and evaluate overall performance of the algorithm. Evaluation images are 135x135 in size.

Results