Samsung R&D Institute China-Beijing (SRC-B) emerged as one of the leading teams at two recent highly prestigious global AI challenges: the New Trends in Image Restoration and Enhancement (NTIRE) workshop at the Conference on Computer Vision and Pattern Recognition (CVPR) and International Workshop on Semantic Evaluation (SemEval) at the Association for Computational Linguistics (ACL). These victories manifest how SRC-B has a strong capability in the semantic understanding and image deblurring fields and the unstoppable spirit of innovation.
NTIRE at CVPR 2021
The NTIRE workshop is held in conjunction with the Computer Vision Foundation (CVF) Open Access. This workshop aims to provide an overview of the new trends and advances in those areas and offer an opportunity for academic and industrial attendees to interact and explore collaborations. The NTIRE 2021 challenge on defocus deblurring using dual pixel images aims to reduce spatially varying defocus blur present in images captured with a shallow depth of field.
△ Ranking of the NTIRE 2021 Defocus Deblurring Challenge Using Dual Pixel Images
SRC-B has achieved the best results for all metrics, including the peak signal-to-noise ratio (PSNR) at 27.80 dB, structural similarity index measure (SSIM) at 0.8484, and inference time at 0.029 s/MP Compared with the second top team, SRC-B’s method is 0.32 dB higher on PSNR and 0.0097 higher on SSIM, while the time consumption is 1/18th of the second team’s.
In the field of blur detection and image deblurring, SRC-B has accumulated vast experience. Before image deblurring, it is vital to determine whether the image is blurred or not. Therefore, the SRC-B team proposed a defocus blur detection algorithm based on deep learning and achieved the State-of-the-Art on open benchmarks and successfully commercialized in Samsung’s products. They have also made impressive achievements in image quality enhancement, in which algorithms, such as image brightness enhancement and image super-resolution, are commercially used in Samsung’s products.
△ SRC-B Team who participated in NTIRE at CVPR 2021
SemEval-2021 at ACL
SRC-B took part in SemEval-2021 Task 4, Reading Comprehension of Abstract Meaning, and secured first place on subtask one and second place on subtask two. In line with this, the related paper was also accepted by the associated conference.
SemEval is an ongoing series of evaluations of computational semantic analysis systems organized by the ACL. It is one of the most influential semantic evaluation competitions. According to Google Scholar, articles published in SemEval are second only to the Association for Computational Linguistics (ACL)/ Empirical Methods in Natural Language Processing (EMNLP)/ North American Chapter of the Association for Computational Linguistics (NAACL) in terms of influence on the field of computational linguistics (also belonging to Samsung Research recommended conferences).
△ SRC-B Team who participated in SemEval-2021 at ACL
Task four is a multiple-choice English Machine Reading Comprehension (MRC) task, aiming to study a machine’s ability to understand abstract concepts in two aspects: subtask one for imperceptible concepts and subtask two for nonspecific concepts. Around 30 teams worldwide participated in the task. The SRC-B team submitted an ELECTRA-based model with task-adaptive pretraining and a multi-head attention multiple-choice classifier on top of the pretrained model and managed to get first place on subtask one and second on subtask two.