Image Fusion for Context Enhancement and Video Surrealism Group No. - 10 Kumar Srijan(200602015) Siddharth Choudhary(200601088) Technical Details • Task is to add context to low contrast night time images using high contrast day time images. • Naïve approaches like simple cutting and pasting of pixels from day time images or averaging will leave artifacts like ghosting , haloing etc. • We will compute gradient of the two images(I1,I2). • Gradient is computed as first difference in both X and Y direction. • Using the gradient of the night image we can compute the importance function(W). Technical details(cont..) • Importance image shows which parts of the night time image contain information. • It can change on application to application. • We are using a binary importance image . • To compute the Importance image, we threshold the gradient of the night time image. • Using the gradients of the day time and night time images and the Importance Image, we can compute the mixed gradient image(G). • Now we have to use this mixed gradient image to construct the final output image. Technical Details Image Reconstruction from gradient field • Approximately an invertibility problem, so solution is not so trivial. • In 2D, a modified gradient vector field G may not be integrable. • So we try to minimize |∇I’− G|, where I’ is the final image. • This can be done by solving the Poisson differential equation 2 ∇ I’= div G. • We will get one equation for each pixel of the output image. • This can be represented graphically as: This shows the system of equations for a 4*4 image Implementation Details • We are using Matlab for coding • We are using the Finite Element Method for solving the system of equations. Difficulties faced • Deciding the Importance image. • Solving the huge system of equations. • Dealing with the boundary conditions for the Poisson equations. Results Obtained • Computed the importance images for night time images. • Found a way to solve Poisson equation. • Got an approximate blending of day and night images to get the context enhanced image with slight color shifts. Input day image Input Night image Importance image Gradient Image in X-direction Gradient Image in X-direction Context Enhanced Image Observations • There are slight color shifts in the output image. Interpretation • Night time image has been visually enhanced. • Blending of the images in gradient domain reduces artifacts like ghosting and haloing. Deliverables completed • Image fusion for context enhancement of static images has been done to a large extent. What remains to be done • Automation of the Computation of the importance image. • Extending this concept to find a way to enhance night time videos.