Hengyu Liu

Academic CV

Computer Vision · GPU Computing · Generative Modeling

Hengyu Liu

Peer-reviewed Publications

  1. Y. Zhang, H. Liu, D. Ding, “A cross-scale framework for low-light image enhancement using spatial–spectral information,” Computers & Electrical Engineering, vol. 106, 108608, 2023. DOI
  2. Y. Zhang, H. Liu, D. Ding, Z. Ma, “Low Light RAW Image Enhancement Using Paired Fast Fourier Convolution and Transformer,” in 2022 IEEE VCIP, 2022. DOI

Published Patent Applications

  • Atmospheric Light Estimation for Dehazing Using Integral-Image Optimization. First-listed inventor · Chinese invention patent application CN117876256A · Published 2024-04-12.
    Algorithmic contribution: Integral/squared-integral image queries replace repeated subregion scans for mean/standard-deviation scoring in quadtree airlight search.
  • Image Dehazing by Refining Atmospheric-Light Masks and Transmission Maps with Depth Information. First-listed inventor · Chinese invention patent application CN117893443A · Published 2024-04-16.
    Algorithmic contribution: Used estimated depth to refine global airlight into a per-pixel mask and adjust transmission, targeting colour casts and weak distant-scene dehazing.

Research Experience

Hangzhou Normal University · Low-Light Image Enhancement

Undergraduate Research · Supervisor: Prof. Dandan Ding · Two second-author papers

Jul 2021 – Jun 2022
  • Low-light enhancement experiments: Suggested FFC, inspired by LaMa, to introduce global context while retaining high-resolution features in RAW-to-RGB enhancement. Contributed architectural modifications and training; compared PSNR, SSIM, and parameter counts.
  • Quality assessment and selective re-enhancement: Proposed quality-based binary selection for further enhancement and tested no-reference IQA methods; drafted CN114565817A (published in 2022; second-listed inventor).

Industry R&D Experience

Hangzhou Edge Exploration Technology Co., Ltd.

Algorithm Engineer (sole algorithm engineer)

Feb 2026 – Jul 2026
  • Sparse-attention symbolic-music modeling: Implemented a symbolic-music model and PyTorch training code from an initial idea of reducing attention computation through musical hierarchy; packing changes reduced recorded peak memory from 67.7 to 38.1 GB (1.24B parameters, 32,768 tokens, 2×H100 80 GB).
  • Structured supervision and training data: Designed MSD (Music Structured Description) as an intermediate generation target to provide structured supervision for controllable generation and targeted editing. Implemented batch MIDI-to-MSD extraction and synthesis of varied requests from each MSD, forming request–MSD–note training triples.
  • Language-model adaptation for music generation: Produced Qwen3-1.7B music-generation checkpoints via vocabulary extension, LoRA continued pretraining and supervised fine-tuning.

ePRO Software (on-site at IM Motors)

Algorithm R&D (on-site contractor)

Apr 2024 – Dec 2024
  • Vision deployment and multi-task evaluation: Deployed quantized YOLOv5 on Qualcomm SA8295 via ONNX/QNN with C++ post-processing; produced a UniAD reproduction and multi-task evaluation setup on nuScenes mini.

ArcVideo Technology Co., Ltd.

Image Algorithm Engineer (internship through Jun 2023)

Feb 2023 – Jan 2024
  • CPU-to-CUDA dehazing acceleration: Reimplemented CPU dehazing in CUDA reusing regional statistics via integral and squared-integral images, with downsampled transmission estimation/guided filtering, buffer reuse and streams. Historical tests with 1080p input recorded approximately 24 ms on i9-9980XE versus 8.5 ms on RTX 3070 Ti.
  • Depth-informed dehazing research: Compared 10+ methods on real failure cases; designed and implemented depth-informed airlight and transmission adjustments. Improved sky colour casts in selected scenes at approximately 9% additional processing time in the tested version; boundary artifacts remained.
  • Motion-feature fusion and interpolation experiments: Modified motion-feature fusion and multi-scale flow connections in an existing network; conducted training and failure-case comparisons. Selected cases showed fewer small-object disappearances, with residual flicker.
  • Edge inference integration: Adapted vision models to Jetson NX/RK3588 using TensorRT plugins, CUDA post-processing, RKNN conversion and image preprocessing.

Selected Open-source Work

  • Guided Filter using CUDA: Standalone full-resolution GPU guided filtering. Published measurements: approximately 9.8 ms / 1080p image on RTX 4090, including allocation and memory copies, excluding the first warm-up run.
  • Integral Image using CUDA: Extended NVIDIA’s shuffle-scan sample for 16-bit input and squared-integral images, using warp-level prefix sums and wider accumulators.
  • MAA Skill Ready (2025): Fine-tuned MobileNetV4 for three skill states, including a cancellable state; reported 99.7% accuracy on 3,000 test images (~16,000 training images). Updated C++ inference integration and training code; maintained and updated ONNX tools, with continued failure-case data expansion and retraining.
  • MAA OCR (2023): Fine-tuned a game-domain Chinese PP-OCRv3 recognizer (MAA #5895).
  • MAA OCR (2026): Contributed to MAA’s recognition-model upgrade from PP-OCRv3 to v6 through ASCII-specific fine-tuning and dictionary-label fixes (MAA #18812, merged into feat/ppocr); implemented a unified pipeline for data generation, training and export (MaaAI #15, under review).

Technical Skills

  • Programming: Python, C++, CUDA C/C++, Linux, Git
  • ML / Deployment: PyTorch, ONNX, TensorRT, QNN, RKNN