Fall 2014

Single-sensor multispectral imaging techniques: Which one is the best? (TAKEN)

Synopsis: The main goal of this project is to study and compare the performance of state-of-the-art multispectral acquisition systems. Multispectral imaging systems record the scene reflections in several spectral bands. Current systems either use different sensors or scan the scene in several shots. To overcome the shortcomings of these systems, the research community has recently started to design single-sensor single-shot multispectral acquisition techniques. In IVRG, we are specifically interested in multispectral imaging where four bands (blue, green, red and NIR) are recorded using a single sensor (seehttp://ivrg.epfl.ch/research/infrared for more information).

To accomplish the goals of this project, we will complete the following tasks:

  1. First, we will conduct a thorough survey on current single-sensor multispectral acquisition techniques (see [1], [2], [3] and [4] as few examples of such algorithms).
  2. In the next step, we will select some of the most important algorithms that represent the main trends in this field and implement them.
  3. We will spend some time recording a dataset of RGB and NIR images representing natural scenes (around 100 pairs of images).
  4. After simulating the acquisition process, we will evaluate the performance of the implemented algorithms on the RGB/NIR dataset. To quantify the quality of images, we will use different measures such PSNR (peak signal to noise ratio), SSIM (structural similarity index) [5], perceptual and reference-free metrics [6].
  5. In the final step, we will study the performance of implemented algorithms subjectively. To this end, we will conduct a user study in which the participants are asked to rate the quality of images produced with different algorithms. The goal of this task is to understand which of the objective metrics studied in the previous step better matches the judgement of human observer especially when images suffer from demosaicing artifacts.


[1] M. M. F. Duarte and R. G. Baraniuk. Kronecker compressive sensing. Transactions on Image Processing, 2012.
[2] Y. M. Lu, C. Fredembach, M. Vetterli and S. Süsstrunk. Designing color filter arrays for the joint capture of visible and near-infrared images. ICIP, 2009.
[3] Z. Sadeghipoor Kermani, Y. Lu and S. Süsstrunk. A novel compressive sensing approach to simultaneously acquire color and near-infrared images on a single sensor. ICASSP, 2013.
[4] L. Miao, H. Qi, R. Ramanath and W. E. Snyder. Binary tree-based generic demosaicing algorithm for multispectral filter arrays. Transactions on Image Processing, 2006.
[6] H. Tang, N. Joshi and A. Kapoor. Learning a blind measure of perceptual image quality. CVPR, 2011.

Prerequisites: Basic knowledge of image and signal processing, good MATLAB skills (familiarity with image acquisition and multispectral imaging is preferred, but not required).

Type of work: 30% literature review, 10% image acquisition, 60% MATLAB implementation. 

Number of students: 1.

Level: MS, semester project, Computer Science or Communication Systems.

Supervisor: Zahra Sadeghipoor (zahra.sadeghipoor@epfl.ch).




Historical Document Image Segmentation (TAKEN)

Synopsis: The problem of Document Image Segmentation involves the extraction of information about the layout components of a document page. A typical historical document contains several components, such as text regions, headlines, illustrations and drawings. It is a necessary pre-processing step for any handwriting recognition system, where the goal is to recognize the handwritten words in a document page. Ideally, a document image segmentation algorithm should be able to accurately separate text regions from any other component of the document page so that the input to a recognition algorithm is as noiseless and reliable as possible.

The goals of this project are:

  • To get familiar with the proposed algorithms in [1,2] for the document segmentation problem. 
  • To implement them and compare them in a dataset of historical document pages. 
  • Provide findings about their differences in accuracy and their applicability on the different documents under investigation. 


[1] Mehri et al. Old document image segmentation using the autocorrelation function and multiresolution analysis. Document Recognition and Retrieval XX, San Francisco, USA, 2013.

[2] Cohen et al. Robust Text and Drawing Segmentation Algorithm for Historical Documents, HIP 2013.

Deliverables: The student should provide in the end of the semester a written report related to the work done. The whole implementation of the segmentation algorithms used for generating the results in the report should also be provided.

Prerequisites: Basic knowledge of Image Processing, programming with MATLAB for signal/image processing OR C++ using the OpenCV library.

Type of work: 20% research, 80% implementation.

Number of students: 1.

Level: BS, semester project, Computer Science or Communication Systems.

Supervisor: Nikolaos Arvanitopoulos (nick.arvanitopoulos@epfl.ch).