Biomedical Imaging Group
EPFL   Student Projects: Raphaël Tornay
English only    BIG > Teaching > Student Projects > Raphaël Tornay
Home page
Tutorials and Reviews
Download Algorithms
bulletStudent projects
Restoration of mosaic images for bright field microscopy

Raphaël Tornay
Section Microtechnique, EPFL

Semester project
June 2004

Image mosaicing is a popular way to obtain a wide field of view image of a scene while preserving the details. This is especially true when studying the structure of the neural system. One special feature of the synaptic tree is its huge ratio between the dendrites’ size and the tree’s spread. Thus, it is crucial to have a wide field of view in keeping a high image-resolution. After assembly, the mosaics must be de-convoluted to attenuate the effects due to the defocalisation. However, discontinuities induced by acquisition-system side effects, or lacking images can provoke the de-convolution program to fail.

We present in this work several tools to remove such issues and we show some results of a mosaic being assembled according to these methods.

Due to the very nature of bright field microscopy, the illumination on the sample is not evenly distributed. Assembling images without any correction induces huge seam effects. We propose to dispose of such seams in fitting on the input images a paraboloid and subtracting it from those images. We called this method background suppression. In our experiment, we subtracted to each input image the same paraboloid, which proved to bring the best results.

Figure. On the left, a "degraded Lena". We simulated the artefacts of the mosaicing of the microscope on Lena. The image is divided into 16 cells, of which two are missing. A non-uniform background has been added to each cell to simulate the spherical aberration and the non-uniformity of illumination.
On the right, we show the result after processing. The discontinuities are almost invisible.

Even though the background suppression dramatically improved the quality of the mosaics, some seam effects could eventually be still visible. Thus, we add a second step of mosaic post-processing called seam elimination.Its principle is the subtraction of a function, which cancel the gap between two pixels on each sides of a seam. This function must be smooth to preserve the details. In our experiments, we built this function by calculating the gaps on the seams and making these intensities lessen slowly through a relaxation algorithm. This method was able to dispose of high-frequency seam effects.

A second issue to be taken care of was to fill the missing images by something reasonable for the de-convolution program. We experienced a relaxation method to build missing cells. More precisely, we made an analogy between the intensity of the pixels and a temperature map. The existing pixels had a constant “temperature” and the blanks were filled in iterating the heat-equation. This produces very smooth transition between existing cells and blanks.

The conjugation of the three methods presented in this report resulted in an important improvement of the mosaic image before sending it to de-convolution program.