关于我

Hello everyone, I am CS (Chen Shuai), an algorithm engineer specializing in machine vision, computer graphics, and deep learning. I am passionate about computer science and artificial intelligence, dedicated to applying advanced algorithms to solve real-world challenges.

Tech Stack

• Programming Languages: C/C++, Python, Java, GLSL
• Frameworks & Libraries: TensorFlow, PyTorch, OpenCV, OpenGL, Qt, MNN, NCNN, CMake, ImGui
• Algorithm Domains: Computer Vision, Deep Learning, Generative Models, LLM
• Development Tools: Visual Studio, Visual Studio Code, PyCharm, Jupyter Notebook, Android Studio, Git, Docker

Open Source Projects

• Computer Vision
  • Open-Vocabulary Multi-Label Classification: Open-vocabulary multi-label image classification system.
  • Scene Recognition CNN: Scene recognition system based on CNN for natural environment classification.
  • Deep Learning of Machine Vision: Exploring the application of deep learning in computer vision.
  • Vision-Trasformers Matting: Image matting technology based on Transformer.
  • YOLO Auto Target: YOLO-based target detection for FPS games.
  • GoogLeNet Inception/Alexnet/ResNet/VGG/ViT: Implementation of classic convolutional neural network architectures.
  • MNN with QT: Combining MNN inference framework with Qt for cross-platform deep learning applications.
  • OpenCV Demo: Example code and applications of the OpenCV library.
• LLM & Agent
  • Lover LLaMA: LLaMA-based romantic dialogue model, including dataset, fine-tuning, quantization, merging, and API deployment.
• Android & Windows Applications
  • Android Camera Special Effect by CS: Android beauty camera with face slimming, big eyes, and teeth whitening. Implemented using OpenCV + MNN + NCNN + GLM + OpenGLES3.0.
  • BokehDepthByCS: Windows AI real-time bokeh software. A C++17 CMake OpenGL project deploying models with NCNN and MNN. Uses GLFW, Glew, glm for rendering, and ImGui for GUI.
  • Matting Tool by CS: Deep learning-based portrait segmentation method (10k+ downloads).
  • Interpolation Tool by CS: Deep learning-based image interpolation method.
• OpenGL/Shader
  • 7 Sharpen Methods: Introduction to 7 sharpening methods and their implementation in C++, Python, Shader, and GLSL.
  • OpenGL CMake by CS: One-click CMake build for OpenGL projects, combined with ImGui.
  • OpenGL with QT: Using Qt to build OpenGL projects.
  • OpenGL Colors Filters: OpenGL color filters.

Project Experience

• XX Company: Algorithm Engineer (2022–Present)
  • Video Analysis & Processing for Video Transmission Monitor (Tech Lead) — Oct 2019–Present
    • Description: Real-time remote monitoring of video transmission data on Android/iOS devices. Supports soft/hard decoding for Full/Limited Range, performing high-performance real-time image processing, analysis, and rendering on video streams.
      • High Performance & Low Power: Real-time 60fps processing of NV12/YUYV sources, memory optimization, zero-copy texture rendering, offline rendering, and high-performance overlay of multiple processes.
      • Full Features: Implemented scaling, 3D LUT, grid lines, false color, dual-threshold zebra patterns, sharpening, magnifier, peak focus, waveform, histogram, vectorscope, ToF point cloud rendering, etc.
      • Cross-Platform: Deployed on Android and iOS; released demos for Windows and Linux.
    • Impact: 345k downloads on the Apple Store in the last year (2023.12–2024.12).
  • Live Streaming Beauty Filter (Project Lead) — Nov 2023–Present
    • Description: Real-time beauty processing for live streaming scenarios on Android (Snapdragon 8250) and Windows (GTX1050Ti+). Enhances skin aesthetics, compensates for ISP limitations, and improves stream quality. Optimized AI-to-zero-copy texture conversion, OpenGL thread synchronization, and pre-initialization, achieving <5ms rendering time and stable 30fps streaming.
      • Stickers: Implemented image algorithms and rendering for sticker translation, rotation, scaling, and mirroring; triangulation and rendering of face stickers aligned with facial keypoints.
      • Skin Segmentation: Used a lightweight skin matting model with optimized datasets for different skin tones and facial features. Created optimized Trimap and alpha datasets. Utilized SNPE 2.22 PTQ quantization and DSP backend with parallel inference, reducing latency by 14ms per frame.
      • Proprietary Green Screen Keying: Efficient background removal with precise foreground-background separation. Includes color space conversion (RGB to YUV), distance calculation, erosion, Gaussian blur, and foreground synthesis.
      • Whitening LUT Module: Generated .cube and .3dl files for varying degrees of whitening. Used PBO zero-copy and frame-skipping schemes. Combined AI skin segmentation with traditional methods (Gaussian probability) and Kalman filtering for stability.
      • Skin Smoothing: Implemented OpenGL-based smoothing to avoid harsh effects, ensuring natural skin texture. Solved gamma brightening issues and edge color bleeding using Gaussian and color distance weights.
      • Face Slimming/Eye Enlarging/Teeth Whitening (R&D): Implemented special effects to enhance personalization and engagement.
    • Impact: Officially launched in company camera products and live streaming platforms.
  • Real-time Alpha Matting Model & Deployment (Project Lead) — Nov 2023–Present
    • Description: Fast and stable alpha matting for human skin regions in single or multi-person (up to 5) environments.
      • Model Optimization: Based on MobileNetV2 backbone DeeplabV3+. Changed input from 3480480 to 4(3320*320) (stacked channels) to reduce latency by ~6ms. Replaced resize operators with deconvolution for high-performance low-latency alpha matting.
      • Inter-frame Stability: Applied Kalman filtering on face detection regions.
      • Deployment: Mobile deployment via SNPE PTQ quantization on DSP; PC deployment via NCNN. Used OpenCL for pre-processing, achieving 11ms inference time while maintaining precision.
  • AI Bokeh for Live Streaming (Tech Lead) — Nov 2023–Present
    • Description: AI Depth Estimation Model and DSLR-like Bokeh Rendering to simulate shallow depth of field.
      • Depth Estimation: Implemented DPT V2 monocular depth estimation (DataLoader & training code). Provided lightweight PyDNet2 model, converted to MNN/NCNN, with C++ streaming code.
      • Focus Distance Stability: Kalman filtering on face detection regions.
      • Depth Map Stability: Frame-based Kalman filtering for stable bokeh simulation.
      • DSLR Rendering: Combined AI depth maps, portrait masks, and DP inputs. Simulated bokeh effects for different lens F-stops and focal lengths using OpenGL, offering more realistic results than competitors.
  • Video Denoising (Tech Lead) — Nov 2023–Present
    • Description: NLMeans-based video denoising to reduce luminance and chroma noise from sensors.
    • Performance: GPU-accelerated OpenGL implementation achieving 60fps denoising.
  • Open-Vocabulary Multi-Label Image Classification System (Tech Lead) — Nov 2025–Apr 2026 (Note: Please verify these dates as they are in the future)
    • Description: Zero-shot open-vocabulary multi-label classification system deployed in AI Smart Editing Products. Supports automated scene tag recognition for video materials, handling 5000+ categories without retraining.
      • Two-Stage Training Strategy: Stage 1 freezes CLIP backbone for knowledge distillation; Stage 2 introduces Prompt Tuning for joint optimization, significantly improving zero-shot generalization.
      • Backbone Adaptation: Successfully adapted DINOv3 and other backbones. Designed a projection layer to map 768-dim visual features to 512-dim semantic space for cross-modal alignment.
      • Multi-Objective Optimization: Combined Cross-Entropy, Contrastive, and Ranking losses, improving recall for rare classes by 15%.
      • Distributed Training & Optimization: Implemented single/multi-node multi-GPU parallel training with mixed precision, improving training efficiency by 3.5x.
      • Performance: Achieved zero-shot mAP of 41.2% on NUS-WIDE (surpassing baseline BiAM 37.6% and AAAI 2023 Oral paper performance). Training throughput increased by 2.9x, and model parameters reduced by 75.5%.
      • Deployment: Model quantized and deployed via ONNX in AI smart editing products.

Papers and Patents

1. Radar Reflectivity and Meteorological Factors Merging‐Based Precipitation Estimation Neural Network[J]. Earth and Space Science, 2021, 8(10). pdf
2. Offline Single-Polarization Radar Quantitative Precipitation Estimation Based on a Spatiotemporal Deep Fusion Model[J]. Advances in Meteorology, 2021.
3. A Cloud-Removal Method for Snow Product Based on Denoising Autoencoder Neural Network[J]. Journal of Nanjing University of Information Science and Technology (Natural Science Edition), 2023, 15(02).
4. Cloud-Removal Algorithm and Application Research for Snow Product at Basin Scale[D]. 2022. DOI:10.27248/d.cnki.gnjqc.2022.001085.
5. A Vehicle Vibration Noise Detection Alarm Device[P]. Jiangsu Province: CN202120495866.X, 2022-03-15.

Education

• Nanjing University of Information Science and Technology: Master’s Degree (2019–2022)
• Nanjing University of Information Science and Technology: Bachelor’s Degree (2015–2019)

Personal Blog

• My Personal Blog: Sharing technical articles, project experiences, and industry insights.

Social Media

• Zhihu: https://www.zhihu.com/people/ou-nuo-40
• CSDN: https://blog.csdn.net/sCs12321/

Contact Information

• Email: cs1179593516@gmail.com
• Github: https://github.com/CSsaan

I am enthusiastic about exploring new technologies, solving practical problems, and enjoy exchanging and collaborating with peers. If you are interested in my work or have any cooperation intentions, please feel free to contact me. I look forward to exploring the infinite possibilities of computer science with you!