Mastering Image Classification: A Journey with Cakes

What happens when your AI fails so badly it decides 'the cake is a lie'? Discover the practical path to mastering image classification.

#1about 4 minutes

The origin of the "Is it Cake?" machine learning project

Inspired by the TV show "Is it Cake?", a personal project was started to build an image classifier using TensorFlow.js.

#2about 3 minutes

Sourcing and preparing the cake and not-cake image data

Data was collected by scraping bakers' websites with Playwright and using the Unsplash API, but this introduced data quality issues like logos and ambiguous images.

#3about 3 minutes

Evaluating the pre-trained MobileNet image classification model

The pre-trained MobileNet model in TensorFlow.js was tested but performed poorly, often misclassifying cakes as candles or bakeries.

#4about 2 minutes

Using the Coco-SSD model for object detection

The Coco-SSD object detection model performed better than MobileNet but still made significant errors, like identifying a cake as a person.

#5about 6 minutes

Building a custom convolutional neural network from scratch

An attempt to build a custom Convolutional Neural Network (CNN) using TensorFlow.js sequential models resulted in failure, with the model classifying every image as "not cake".

#6about 2 minutes

Improving model accuracy with transfer learning

Transfer learning was used by taking the feature extraction layers of MobileNet and adding a custom classification head, which significantly improved the model's performance.

#7about 4 minutes