Miniatures

UC Berkeley CS-194 Final Project

Helen Che

December 8, 2017

In this project, I took some stock photos of the internet (all stock photos courtesy of Pexels) and manipulated them to look like little toys by mimicing a tilt-shift. I did this by building a gaussian pyramid + alpha mask blending. Below are the steps I took to achieve a toy-like appearance with an airplanes photo (compared directly below are the final result and original).

1. Bumping saturation

before starting, I decided to increase the saturation to produce a better overall effect for the miniature effect. Enhacing the saturation makes objects in the image look more artifical since the colors are much brighter. To do this, I first converted the image from RGB (red green blue) channels to HSV (hue saturation value) channels. I then bumped the saturation channel by 20% and topped off the values so they wouldn't exceed 1. Below are the original image and saturated images side by side.

2. selecting depth of field and line

My code allows someone to choose two points that will define the depth-of-field line. For this particular image, the red lines below indicate the region which remains in focus. The best effect happened with a DOF of 120 pixels above and below the center line.

3. Gaussian stack and mask blurring

To achieve the blurred effects, I used a guassian pyramid on each channel of the image and then combined the channels. I first seperated the image into a gaussian pyramid with 4 layers. I then divided the images into regions using masks starting from the outer region (furthest from the DOF region). I gaussian blurred the mask with a constant sigma=8. The most blurry Guassian layers were masked in the outer most layers. To combine all the layers, I masked and then recursively overlayed the layers on top of the un-blurred image in order from least to most blurry (below, from right to left):

(pseudocode):

def make_img(image, num_layers=5):

if num_layers=0: return image

else:

gaussian_mask = gaussian_blur(mask(num_layer), sigma=8)

gaussian_img = gaussian_blur(image, sigma=current_layer_sigma)

up_level_gaussian = make_img(image, num_layers-1)

final_img = gaussian_img * gaussian_mask + (1 - gaussian_mask) * up_level_gaussian

return final_img