When taking a picture of outside scenery, image degradation can be caused due to several reasons. One of these reasons is fog - a layer of clouds residing between the subject and the camera
Abstract
When taking a picture of outside scenery, image degradation can be caused due to several reasons. One of these reasons is fog – a layer of clouds residing between the subject and the camera, blocking or interfering with some of the information on the image. Foggy images are degraded in detail and color fidelity, affecting both human viewing and the ability to use computer vision applications.
This project relays on a physical model, according to which the fog appears different when viewed via different polarizers, which are placed between the viewer and the cloud layer. This model can be exploited, using a Blind Source Separation (BSS) technique, for estimating the original scenery image by using two differently polarized foggy images.
During this project, a special bi-channel camera was used, which is capable of taking two pictures simultaneously with two different polarizers.
Flowchart
Sparsification
- Achieving single active source done via sparsification
- For sparse sources, single source active probability is a non decreasing function
- A ratio for one active source can be found
- STFT on 1D, Haar on 2D
Results
Conclusions
- Ran Kaftory’s algorithm can be applied for foggy images
- The algorithm is independent of image orientation
- The quality of reconstruction is strongly dependent on the compatibility of the sources to the assumed model
- Despite inherent Dual-channel camera advantages, DSLR offered better results
Further work
- Automatically estimating the process parameters
- Robustness for images with small differences – leaves, cars and so on
- Processing images with arbitrary size, not just power of 2
- Reconstruction in real time, for video processing