Projects

Long-term tracking of individual crimson snapper in aquaculture is difficult: inter-individual appearance differences are subtle, underwater image quality fluctuates, and in real deployment new and unregistered individuals appear continuously — so a closed-set recognition assumption does not hold.

Build an open-world recognition method on vision-language priors that both identifies known identities and rejects unknown ones, while compressing the model to a deployable size.

Designed a Compact256 projection layer on the CLIP-ReID backbone to compress feature dimensionality by 5×; proposed a Quality-Aware Identity Prototype Mechanism (QACM) that folds image-quality estimation into identity prototype modelling, unifying known-identity recognition and unknown-identity rejection; built a crimson snapper individual dataset covering 92 identities and over ten thousand images, then ran both closed-set and deployment-level open-set evaluations.

This project delivers a single-fish video runtime system for open-world fish individual re-identification. Centred on track-level identity lifecycle management, five core mechanisms — including Fast Warm-up, post-confirmation identity retention, and dynamic confidence updating — solve stable recognition under track breaks, frame-quality fluctuation, and identity conflicts.

Dual-modality crimson snapper recognition research needs labels and multi-view images captured at the same instant, but manual shooting and annotation cannot guarantee a strict ID-to-image correspondence — data consistency was poor and throughput low.

Own the design and delivery of the complete device: RFID-triggered multi-camera synchronized capture, with data automatically archived per individual ID.

Designed a cross-wall inverted-U UPVC fish passage integrating a 134.2 kHz RFID reader unit and three global-shutter vision units; wrote a multi-threaded Python acquisition program delivering RFID-triggered recording, three-camera synchronization, and automatic per-ID data archiving; resolved the multi-stream bandwidth bottleneck via USB 3.2 and documented a complete engineering build-and-operation standard.

Three 25 fps video streams are captured stably, and acquisition-plus-annotation efficiency improved by 80% over the manual workflow. A reusable engineering build-and-operation standard now supports downstream dataset production.

Research projects, model experiment results, academic papers and resume information in computer vision lacked a single online showcase — interviewers could only learn about the research and engineering work from a static resume.

Independently build a Next.js-based personal technical portfolio: project display modules, fully responsive pages, and a public deployment that interviewers can browse online.

Built the front end and content structure with Next.js, TypeScript and Tailwind CSS; managed the source with Git and wired an automated Vercel deployment pipeline; implemented project display modules and an image lightbox gallery, optimised first-screen performance with SSG static pre-rendering and CDN caching, and maintained technical articles in MDX.

Delivered the public deployment with full-site SSG static pre-rendering and CDN caching, improving first-screen loading performance by over 60%; built a research showcase that visualises experiment data, flowcharts and key metrics of the fish individual re-identification projects; implemented a GitHub + Vercel automated deployment pipeline that releases after each commit, cutting the launch cycle to 2 minutes; and now serves as the online job portfolio in place of a traditional static resume.

Individual tagging of crimson snapper lacked a unified process standard: survival and tag-retention rates varied markedly across fish sizes and between chip implantation and external marking, making results hard to compare.

Screen applicable tagging methods through controlled multi-method experiments, establish standardized operating and post-operative care procedures, and produce selection guidance by fish size.

Designed controlled multi-method experiments and ran the full workflow on over one hundred fish across different size classes; established a standardized chip implantation and post-operative care procedure; compared dorsal injection marking and other approaches for survival and retention across size classes, and consolidated the results into selection criteria.

Dorsal injection marking survival for adult fish improved to 80%; a tagging-method selection guide by fish size class was produced, providing a stable marking basis for subsequent individual-recognition dataset construction.