Session: 16-02-01: Poster Session: NSF Research Experience for Undergraduates (REU)

Paper Number: 77302

Start Time: Wednesday, 02:25 PM

77302 - Designing a Degradation Model Based on Prior Hardware Knowledge for Blind Image Super Resolution 

Recent research on super resolution (SR) has been developed using deep convolutional neural networks (DCNN). Many DCNN architectures have proven to offer accurate models for training datasets.  However, when introducing these models to images outside of the training datasets, they exhibit less favorable performance. This is because the low-resolution images used for training are artificially created from high-resolution images. Therefore, blind super resolution (BSR) has been a highly researched field in the past few years. BSR is an effort to create low-resolution images that are more similar to real-world low-resolution images, so SR models can be trained to yield favorable results for broader testing data.

In this project, we propose a DCNN that generates low-resolution images from high-resolution images, using the enhanced residual blocks from (1) with the addition of our custom layer of weighted downsampling kernels. The proposed solution to the BSR problem was motivated by the need of a robust, accurate way of assessing the quality of roll to roll (R2R) printing machines, where small circuits are printed on flexible substrates. The application of SR to the R2R flexible electronics printing process is essential because of the necessity of a 20× magnification factor for optical imaging. However, applying such a high magnification factor lens is limiting in R2R printing scenarios. For example, a 20× lens has very narrow depth of focus that will cause blurriness in the imaging. SR helps reduce the need for such high-level optics by enabling the use of lower-level optics, for example 5×, where motion is in the tolerance of a wider depth of focus. In this way, the integration of a SR algorithm with a low-resolution lens can represent an accurate high-resolution image. This also helps reduce the cost of the optical devices. In addition, the pattern resolution limit and the fundamental optical diffraction limit of the hardware can be surpassed via SR.

To create low-resolution images, our proposed architecture takes high-resolution images as input. The high-resolution image first goes through enhanced residual blocks, which transfers the information into a high-dimensional representation. Then, several weighted blurring and downsampling kernels are implemented to filter each dimension of the representation. The kernels themselves are fixed and chosen based on prior knowledge from the imaging hardware. We chose bicubic with anti-aliasing, isotropic Gaussian, and anisotropic Gaussian kernels. Last, the filtered high-dimensional representation is decoded back into the low-resolution image.

We initially trained our model using the SupER dataset of pictures from (2), each with different amounts of contrasts for variability. We chose this dataset because of the authors’ method of establishing ground truth pictures, which considers sensor noise. The trained model was tested on the RealSR dataset taken from (3). We then plan to train and test our final model with our own self-made datasets. The proposed method of BSR has displayed promising results for generating accurate low-resolution images, in comparison with bicubic downsampling.

 

1.         B. Lim, S. Son, H. Kim, S. Nah, K. Mu Lee, in Proceedings of the IEEE conference on computer vision and pattern recognition workshops (2017), pp. 136–144.

2.         T. Köhler et al., IEEE Trans. Pattern Anal. Mach. Intell. 42, 2944–2959 (2019).

3.         J. Cai, H. Zeng, H. Yong, Z. Cao, L. Zhang, in Proceedings of the IEEE/CVF International Conference on Computer Vision (2019), pp. 3086–3095.

 

 

Presenting Author: Johnathan Czernik University of Massachusetts Amherst

Authors:

Johnathan Czernik University of Massachusetts Amherst
Rui Ma University of Massachusetts, Amherst
Xian Du University of Massachusetts Amherst