Designed a few-shot defect image generation method based on vision foundation models and flow matching, addressing the scarcity of defect samples in industrial anomaly detection. Formulated defect synthesis as a context-aware local editing task: a dual-branch defect feature extraction module captures fine-grained defect semantics, while a FLUX.1 Kontext-based generation pipeline combines inpaint-aware flow matching, defect feature matching, and normal region preservation constraints. A "Generate-Select-Refine" three-stage mask generation flow was proposed to improve defect-to-image alignment quality. Responsible for method design, model training, and experimental analysis.

Investigated memory optimization for LLM fine-tuning in resource-constrained environments. Built a GPU memory estimation model for Decoder-only Transformers, quantifying various memory demands. Systematically surveyed LoRA, QLoRA, activation recomputation, and parameter offloading strategies from both algorithmic and system perspectives. Conducted comparative experiments with 5 fine-tuning strategies on LLaMA3.2-3B using the GSM8K dataset, evaluating memory cost, training time, and convergence quality.

Proposed a hierarchical instruction generation model based on a two-scale graph Transformer (DUET) and a large language model to address data scarcity and low instruction quality in UAV vision-language navigation. The method uses heuristic search to generate paths, extracts multi-modal features via ViT-B/16 and BERT with a dynamic merging strategy, and leverages DUET's dual-scale visual representations with LLM contextual reasoning to generate high-quality navigation instructions. Filtered by BLEU metrics, the augmented data significantly improves instruction conciseness and key action density.