Computer Vision Research — 2025 — 2026
Malaysian Kuih Recognition
“A classifier small enough to run on a phone, and a dataset that did not exist before.”
Role
First author — dataset construction, model training, system integration
Context
IEEE Access · IJ-AI · ICERCS 2025 · Best Project, COMMAX
Focus
Peak accuracy
Inference time
Images curated
Classes
Four papers and a built system, all resting on one thing that had to come first: a hybrid dataset of 1,227 images across 11 classes of East Coast Malaysian kuih. The models are lightweight by requirement, not preference — the target is a phone in a kitchen, not a GPU in a lab.
The problem
Malaysian kuih is not represented in general food-recognition datasets, and the classes are genuinely hard: many varieties differ by texture and internal composition rather than silhouette or colour, so a model has to discriminate on fine detail. Nutritional data adds a second problem — a verified calorie lookup is accurate but says nothing about what the food is, while a language model is fluent about cultural context and unreliable about numbers.
What I built
A hybrid dataset of 1,227 images was assembled and used to train and compare CNN architectures under a shared evaluation protocol: MobileNet-V2 first at 95.8%, ResNet50 at 95.41%, and a tuned lightweight MobileNetV2 configuration reaching 98.17% at 58 ms inference. The nutrition problem was resolved by refusing to pick a side: a KKM-sourced lookup supplies verified calorie figures, and Gemini 2.5 Flash supplies cultural and preparation context, presented as two clearly distinct readings so the user always knows which number is authoritative.
How it works
Building the dataset
1,227 images across 11 East Coast kuih classes, assembled as a hybrid of captured and sourced imagery. This was the bulk of the work and the precondition for everything after it.
Comparative architecture study
MobileNet-V2, ResNet50 and a tuned lightweight MobileNetV2 trained under one protocol so the numbers are actually comparable — the point being accuracy per millisecond, not accuracy alone.
Optimising for edge inference
The target is a phone, so 58 ms inference is a requirement rather than a bonus. Architecture selection was driven by that ceiling.
Dual-reading nutrition interface
KKM lookup for verified calories, Gemini 2.5 Flash for cultural and preparation knowledge, kept visually separate. A generative model should not be the source of a number a user might act on.
Outcome
Four papers with Rahimi as first author — IEEE Access, IAES IJ-AI, ICERCS 2025, and a ResNet50 study — plus a working prototype that won Best Project at COMMAX. The dataset is the durable contribution: it is the part that did not exist before and the part the later work depends on.
Stack