WorkMalaysian Kuih Recognition
AI / Computer Vision · 2026
Malaysian Kuih Recognition
“A classifier small enough to run on a phone, trained on a dataset that did not exist before.”
Services
AI Solutions
Tech
PythonTensorFlowCNNMobileNetV2Edge InferenceGenerative AI
Built by
Full case study →
Rahimi Rusyuddin
First author, dataset construction, model training, system integration
Personal research project. Four papers as first author across IEEE Access, IAES IJ-AI and ICERCS 2025. Best Project at COMMAX.
98.17%
Peak accuracy
58ms
Inference time
1227
Images curated
11
Kuih classes
A lightweight image classifier that identifies 11 varieties of East Coast Malaysian kuih and returns a calorie reading, resting on a 1,227-image dataset built from scratch because no such dataset existed.
The Challenge
Malaysian kuih does not appear 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 the model has to discriminate on fine detail. Nutrition 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.
Our Solution
Three CNN architectures were trained under one shared evaluation protocol so the numbers were actually comparable: MobileNet-V2, ResNet50, and a tuned lightweight MobileNetV2. Selection was driven by accuracy per millisecond rather than accuracy alone, because the deployment target is a phone in a kitchen, not a GPU in a lab. The nutrition question was settled by refusing to pick a side: a KKM-sourced lookup supplies the verified calorie figures and Gemini 2.5 Flash supplies cultural and preparation context, kept visually separate so the reader always knows which number is authoritative.
The Outcome
The tuned MobileNetV2 reached 98.17% accuracy at 58 ms inference, ahead of MobileNet-V2 at 95.8% and ResNet50 at 95.41%. Four papers followed with Rahimi as first author, and the prototype took Best Project at COMMAX. The dataset is the durable contribution: it is the part that did not exist before, and the part every later result rests on.
Full Tech Stack
PythonTensorFlowMobileNetV2ResNet50Gemini 2.5 FlashMySQL