Deghosting

• Avinash Prabhu - 2018102027
• Mallika Subramanian - 2018101041
• Fiza Husain - 2018101035
• Tanvi Karandikar - 2018101059

Implementation of the paper : https://www.cv-foundation.org/openaccess/content_cvpr_2015/papers/Shih_Reflection_Removal_Using_2015_CVPR_paper.pdf

Matlab Toolkits Required

• Image Processing Toolbox
• Partial Differntiation Toolbox
• Signal Processing Toolbox

How to run the code

cd src/
deghost(image_path, label_x, label_y, dk_thresh, ck_thresh)

For an image of size 300x200, it takes <25 minutes to run on 30 CPUs of 2GB RAM each. It is not recommended to run this code for large images on your local system as it may cause your laptop to hang and crash.

Inputs

• image_path: Relative(wrt src/) or absolute path of the image to be saved. Image should be in png format.
• label_x: Direction of horizontal shift (see below)
• label_y: Direction of vertical shift (see below)
• dk_thresh: Threshold value to be used to select which candidates from the autocorrelation should be considered for dk shift.
• ck_thresh: Threshold value to be used to select which candidates are considered as patches with 'high correlation'

Selecting label_x and label_y

In order to improve output accuracy, the user can give an input in which they specify the direction of the shift. We use shift of the second reflection wrt to the first. User is required to input their estimate of the shift direction.
First reflection is the brighter/sharper one.

Second reflection is to the <label_x> of the first reflection.

Horizontal Direction	Label x
Left	l
Right	r
Unsure	-

Second reflection is to the <label_y> of the first reflection.

Vertical Direction	Label y
Down	d
Up	u
Unsure	-

"-" can be inputted when the shift direction is not obvious.

Choosing the thresholds ck_thresh and dk_thresh

ck_thresh and dk_thresh will differ from image to image, since they depend on factors such as lighting, the thickness of the glass, etc.

In general:
dk_thresh can be taken in the range 0 - 100
ck_thresh can be taken in the range 0.5 - 0.9

Outputs

The algorithm outputs the separated reflection and transmission layers as separate images. Outputs are saved in the outputs/ directory. The outputs are:

• <input_image_name>_input.png : The input image
• <input_image_name>_reflection.png : The extracted reflection image
• <input_image_name>_transmission.png : The extracted transmission image

Example

In this image we can clearly see that the rose is part of the reflected portion, by ghosting cues.
The fainter rose (second reflection) is to the top and left of the darker rose (first reflection).
Hence we pass label_x as "l" and label_y as "u".

Function call:

deghost("../images/prova_2.png", "l", "u", 0, 0.85)