- Visual anomaly detection
- Foundation models for industrial anomaly detection
- Parameter-efficient adaptation and test-time adaptation
- Continual learning
- Time-series anomaly detection with foundation models
- Sep 2025 - Present: MSc, Computer Science and Technology, Xinjiang University. Supervisor: Panpan Zheng.
- First-author paper under review (2026) in visual anomaly detection.
- Research spans industrial, logical, and medical anomaly-detection settings using CLIP, DINOv2, SAM, LoRA, adapters, and visual prompts.
- AF-CLIP Reproduced - zero-shot anomaly detection across six industrial benchmarks. Five paper-reported benchmarks match at the paper's one-decimal precision; MVTec-LOCO is included as an additional cross-dataset evaluation.
- APRIL-GAN Reproduced - official zero-shot protocol on MVTec-AD and VisA. Maximum absolute paper delta is 0.5 percentage points across the reported aggregate metrics.
- AnomalyCLIP Reproduced - paper-compatible final-layer evaluation with sanitized logs, aggregate summaries, source provenance, and repository verification.
- WinCLIP Reproduced - reference zero-shot results reproduced on MVTec-AD and VisA, with diagnostic implementation differences documented separately.
These repositories are independent reproduction and evaluation work. Upstream methods, code, checkpoints, and datasets remain attributed to their original authors and licenses.
- Moving anomaly-detection models toward C++ and OpenCV deployment.
- Building an ONNX to TensorRT pipeline with INT8 quantization and measured latency comparisons. Results will be published after the measurements are complete.
Python · PyTorch · OpenCV · C/C++ · Hugging Face · Docker · Linux · Git