Authors

Chao Pan 1 , S. M. Hossein Tabatabaei Yazdi 2 , S Kasra Tabatabaei 1,Alvaro G. Hernandez 1 , Charles Schroeder 1 , Olgica Milenkovic 1∗

ABSTRACT

the prohibitively high cost
catastrophic error-propagation
Here we propose the first method for storing quantized images in DNA that uses signal processing and machine learning techniques to deal with error and cost issues without resorting to the use of redundant oligos or rewriting.
Our methods rely on decoupling the RGB channels of images, performing specialized quantization and compression on the individual color channels, and using new discoloration detection and image inpainting techniques.
We demonstrate the performance of our approach experimentally on a collection of movie posters stored in DNA.

nonvolatile

image.png

in a nonvolatile fashion

one may not be able to cover all sub-strings of the user-defined string.

An example illustrating the image corruptions caused by erroneous/missing oligos is shown in Figure 3.

image.png
Small blocks with the wrong color can be easily observed visually, and they are a consequence of only 10 missing oligos.
To correct the discolorations automatically, we propose a three-part image processing procedure.
The first step consists in detecting the locations with discol-
orations, masking the regions with discolorations and subse-
quently treating them as missing pixels.
The second step involves using deep learning techniques to inpaint the missing pixels.
The third step involves smoothing the image to reduce both blocking effects caused by aggressive quantization and the mismatched inpainted pixels.

To the best of our knowledge, detecting arbitrarily shaped discolorations is a difficult problem in computer vision that has not been successfully addressed for classical image processing systems.
This is due to the fact that discolored pixels usually have
simultaneous distortions in all three color channels of possi-
bly different degrees.
However, detecting discolorations in DNA-encoded images is possible since with high probability,only one of the three color channels will be corrupted due to independent encoding of the RGB components.
Figure 4 illustrates this fact, as erroneous pixels in different channels do not overlap.

image.png
Within the correct color channels, pixels have neighbors of similar level, while within the erroneous channel, pixels have values that differ significantly from those of their neighbors.
Figures 5 (a)(b)(c) illustrates that pixels with the smallest t = 15 frequencies in the difference vectors indeed correspond to almost all erroneous regions in the red channel.

image.png
The results of our detection scheme are depicted in Figure 5(d)(e), for t = 18.
Note that the whitened out regions are treated as missing data, and filled in using inpainting techniques.

image.png

Image inpainting, or image completion, is a method
for filling out missing regions in an image.
There exist several methods for image inpainting currently in use, including diffusion-based, patch-based [16] and deep learning approaches [10].
The former two methods use local or non-local information only within the target image itself which leads to poor performance when trying to recover complex details in large images.
On the other hand, deep-learning methods such as EdgeConnect [10] combine edges in the missing regions with color and texture information from the remainder of the image to fill in the missing pixels.
Since the encoded movie posters have obvious edge structures, we inpainted the images using EdgeConnect with the result shown in Figure 6(a).

image.png

Although the problem of discoloration may be addressed through inpainting, the reconstructed images
still suffer from mismatched inpaints and blocking effect
caused by quantization.
To further improve the image quality we perform smoothing through bilateral filtering [11] that tends to preserve the edges structures.

image.png
The smoothing equations read as:

image.png
The filter performs Gaussian blurring on background regions but respects edge boundaries in the image.
The result of smoothing with $σ^{2}_d = σ^2_r = 45$ and $Ω$ of the form of a 9 × 9 square is shown in Figure 6(b), and no obvious discolorations are detectable.
Furthermore, in order to address other possible impairments, we also used the positions of error blocks obtained from the discoloration detection platfrom to perform adaptive median smoothing around erroneous regions.
The output of this iterative process is illustrated in Figure 6(c)(d).