Jingchao Hou

Image demosaicing is one of the most ubiquitous and performance-critical image processing tasks. However, traditional demosaicing methods use fixed weights to finish the interpolation, while deep learning demosaicing restoration always breaks the image array arrangement rule, and they can't fully use the existing pixel information. To rectify these weaknesses, in this paper, we propose the convolution interpolation block (CIB) to obey the RAW data arrangement rule and the deep demosaicing residual block (DDRB) to repeatedly utilize existing pixel information for demosaicing. Based on the CIB and DDRB, we present a novel two-stage demosaicing model (DTDeMo), including differential interpolation and enhancement processes. Specifically, the interpolation process contains several CIBs and DDRBs with trainable interpolation parameters. Meanwhile, the enhancement process consists of a transformer-based block and a series of DDRBs to enhance the interpolation results. The effectiveness of CIBs, DDRBs, the proposed interpolation process, and the enhancement process is confirmed through an ablation study. A thorough comparison with several methods shows that our DTDeMo outperforms state of the art quantitatively and qualitatively.

Recently, the single image super-resolution (SISR) based on deep learning algorithm has taken more attention from the research community. There are many methods that are developed to solve this task using CNNs methods. However, most of these methods need large computational resources and consume more runtime. Due to the fact that the runtime is essential for some applications, we propose a mixer-based local residual network (MLRN) for lightweight image super-resolution (SR). The idea of the MLRN model is based on mixing channel and spatial features and mixing low and high-frequency information. This is done by designing a mixer local residual block (MLRB) to be the backbone of our model. Moreover, the bilinear up-sampling is utilized to transfer and mix low-frequency information with extracted high-frequency information. Finally, the GELU activation is used in the main model, proving its efficiency for the SR task. The experimental results show the effectiveness of the model against other state-of-the-art lightweight models. Finally, we took part in the Efficient Super-Resolution 2023 Challenge and achieved good results.