Jinhui Chen

• Burst super resolution: Designed and trained various deep learning architectures to reconstruct high-quality images from sets of low-quality inputs. Quantized the trained burst super-resolution model, balancing effectiveness and efficiency for hardware deployment. The final quantized model achieves a PSNR of 45.7 with a latency of 20 ms.
• Neural architecture search: Developed a Neural Architecture Search (NAS) framework for multiple projects targeting edge devices, optimizing model performance within given hardware constraints. The framework achieves up to a 1.34% improvement in top-1 accuracy, along with up to 56.08% latency reduction and 41.72% energy savings.
• Foveated segmentation: Designed a framework that segments, detects, and recognizes objects using eye-tracking data, improving the mIoU on the Cityscapes dataset from 0.37 to 0.52. Unlike most segmentation algorithms that perform well on large and medium-sized objects but struggle with small ones, our approach effectively handles small-sized objects.

Video editing:

• Designed rule-based algorithms to extract individual videos of multiple people from panoramic footage. A key challenge was ensuring the resulting videos were user-friendly—free of flickering and jerky movements.
• Designed and trained a deep learning model to replace the rule-based algorithm, using data generated in the first step.

• Night vision: Led the design of a lightweight neural network architecture based on Retinex algorithms, featuring two components: enhancement and denoising. This resulted in a real-time night vision algorithm optimized for mobile video applications.
• Super resolution: Led the design of a lightweight network for video super-resolution based on the ESPCN (Efficient Sub-Pixel Convolution Neural Network), incorporating an attention mechanism that effectively addresses the trade-off between over-fitting and generalization.

• Multi-Human parsing: Trained a model using off-the-shelf algorithms to segment human images into 59 categories, including background, body parts, and fashion accessories. Deployed the model to the cloud.

• Worked in the Department of Information and Computing Science, responsible for teaching the Algorithm and Machine Learning courses.

• LCD anomaly detection: Designed an algorithm to detect anomalies in LCD images by mimicking the zoom-in and zoom-out process, enabling the detection of even the slightest defects. This approach improved overall accuracy from 91% to 98%, and increased Mura detection accuracy—a specific type of LCD anomaly—from 43% to 100%.
• Image quality evaluation: Designed an algorithm for online image quality evaluation in credit card applications, combining Fourier transform and Gaussian filtering to significantly reduce manual effort.

• Advertising deployment system: Led the development and personally designed a machine learning model to maximize target audience reach for commercial advertisements across various media platforms, including streaming services, television, and billboards. The advertising strategies generated by this system significantly increased exposure to target audiences and contributed to improved sales performance.
• Evaluation system: Led the development of an evaluation system to report advertising effectiveness, significantly reducing workload.