Research Projects

• Research on classification and enhancement of dark-light images based on data enhancement

  01/2024 - Present

  • Migrate dark light data to normal open source data style.
  • Enhance reasoning speed through some methods.
  • Enhance images through some methods (such as attention mechanism).

• Advancing Out-of-Distribution Detection through Data Purification and Dynamic Activation Function Design

  7/2023 - 12/2023

  • We present the OOD-R dataset, an innovative amalgamation of existing open-source datasets, distinguished by its low noise level. This rectified dataset, through strategic noise filtering, offers enhanced data quality for OOD detection, providing clearer and more reliable samples for research and model development.
  • We have also introduced the ActFun activation structure, which substitutes traditional ReLU with the expectation version of ReLU in various networks. This change significantly boosts OOD detection’s specificity and accuracy. Notably, ActFun has shown a considerable improvement in evaluation methods, marked by up to 18.42% increase in AUROC and a minimum of 16.93% decrease in FPR95, underscoring the importance of precise hyperparameter calibration in optimizing OOD detection.
  • Our research has examined the impact of the hyperparameter $\beta$ on different OOD detection algorithms. We found a strong correlation between this parameter and each method’s performance, highlighting the need for accurate hyperparameter tuning, especially when modifying activation functions, to enhance OOD detection effectiveness.

• Applying Ming furniture features to modern furniture design using deep learning

  3/2023 - 8/2023

  • By collecting and filtering existing physical images of Ming-style furniture, using a generative adversarial network algorithm (DCGAN) for image recognition and feature extraction, and generating modern furniture designs.

• Research and Optimizing Implementation of a New Stencil Parallel Algorithm

  07/2021 - 06/2022

  • A blocking algorithm is proposed for Gauss-Seidel red-black sorting that is suitable for any problem size, block size, and block starting position. It has lightweight, concise loop conditions and better multi-core inter-core parallelism.
  • Design a new data layout scheme that separates the spatial dimensions and grid points in the data space, and designs a corresponding vectorization scheme to reduce redundant calculations and eliminate the spatial data conflict problem of input data in the vectorization of Stencil calculations.
  • When multi-core parallel is used, the one-dimensional to three-dimensional performance of this algorithm is improved by 3.23 times, 4.15 times, and 2.35 times respectively. 24 cores are 12.94 times, 14.56 times, and 21.09 times faster than serial in one, two, and three dimensions respectively.

• Designing an Emotion Design Approach with Perceptual Engineering and Deep Convolutional Generative Adversarial Networks in Social Robotics

  03/2020 - 08/2020

  • A perceptual engineering-based approach to emotional design is proposed using deep convolutional generative adversarial networks.