Universal SDR-to-HDR Conversion via Bidirectional Degradation Source Prompt Learning
Research on universal SDR-to-HDR conversion and inverse tone mapping.
ABOUT
I'm an undergraduate student in Electronic Information Engineering at Shenzhen University.
My current interests lie at the intersection of multimodal AI, computer vision, computational imaging, and increasingly, large language model systems.
I enjoy projects where I can understand the whole pipeline: preparing data, implementing models, training them, evaluating their behavior, and figuring out why something works — or why it unexpectedly doesn't.
My previous work has focused mainly on image and video problems, including SDR-to-HDR conversion, video quality assessment, and image captioning. More recently, I've been exploring retrieval, language models, and the engineering behind modern LLM systems.
INTERESTS
Connecting visual information with language and other modalities.
Learning representations for visual perception, restoration, generation, and quality assessment.
Learning-based image enhancement, reconstruction, and SDR-to-HDR conversion.
Retrieval-augmented generation, language model architectures, lightweight training, and inference.
RESEARCH
Research on universal SDR-to-HDR conversion and inverse tone mapping.
PROJECTS
Deep-learning-based SDR-to-HDR conversion experiments, including model reproduction, training, evaluation, and investigation of abnormal luminance behavior.
Image captioning experiments with ResNet / ConvNeXt encoders, Attention-LSTM and Transformer decoders, beam search, and CIDEr-oriented SCST optimization.
Reproduction and experimentation with KSVQE, including multi-view preprocessing, distortion analysis, cross-dataset evaluation, and inference.
A lightweight RAG laboratory for academic documents, comparing sparse, dense, hybrid retrieval, reranking, citation, and retrieval evaluation.
A small decoder-only language model implemented in PyTorch, exploring pretraining, autoregressive generation, and LoRA instruction tuning.
EXPERIENCE
Worked on a research project involving model reproduction, dataset preparation, training, evaluation, and analysis of SDR-to-HDR conversion behavior.
Built and optimized data collection and cleaning pipelines for domain knowledge used in an agricultural large language model.
TOOLBOX
Python · C · C++
PyTorch · CNN · Transformer · ResNet · ConvNeXt · ViT · Swin Transformer · CLIP · Attention-LSTM
OpenCV · Image Processing · Video Processing
RAG · FAISS · Embeddings · LoRA
Requests · aiohttp · Selenium
Git · Linux · CUDA
I don't try to collect technologies for the sake of a longer list. Most things here come from projects I've actually worked on or am currently exploring.