- Designed a spatio-frequency fusion strategy with change-guided attention to better capture meaningful landscape transformations.
- Integrated CBAM into the decoder to suppress noise and sharpen change boundaries.
- Improved minority-class sensitivity with Dice Loss and loss functions tailored to changed-versus-unchanged separation.
Applied Research
Projects
Remote sensing Heritage restoration Demographic analytics INR theory
Portfolio
Research and engineering projects
The timeline below focuses on the core problem, technical contribution, and research direction of each project.
- Developed CoRe-INR using axis-wise coordinate scaling and confidence-based fusion; analysed spectral support and convergence through frequency-domain and neural tangent kernel analysis.
- Proposed SSA-guided branch specialisation with explicit component supervision, PSO-initialised learnable scaling, and gated additive fusion for image and audio representation.
- Evaluated CoRe-INR on image, audio, and 3D representation, alongside denoising, super-resolution, and inpainting.
- Contributing to an NVIDIA-funded heritage restoration project centered on generative AI for cultural preservation.
- Developing diffusion-based and implicit neural representation pipelines for restoring degraded historical photographs and documents.
- Implementing INR-driven denoising workflows that remove blemishes, noise, and structural artifacts in heritage imagery.
- Developed PoPDivergence, a KL-divergence-based scalar to compare national population pyramids against an optimized reference.
- Introduced PoPStat to quantify demographic influence on disease-specific mortality across 371 global diseases.
- Benchmarked the method against HDI, GDP, median age, and life expectancy, showing stronger explanatory power for many diseases.
- Developed a preprocessing method that corrects large illumination-induced scale distortions before unmixing.
- Restored the simplex geometry assumed by the Linear Mixing Model to improve downstream abundance estimation.
- Introduced a perspective-projection-based scale-correction algorithm that reduced abundance estimation errors by roughly 50 percent across diverse unmixing methods.