This is a simple notebook that details the steps in creating a CNN based classifer to recognise different types of Singaporean food. The overall accuracy was between 0.75 and 0.80 over 50 epochs.
The approach taken was to use transfer learning (using InceptionV3) to train 16000+ different images across 17 classes of food types.
Two observations are that the model does not differentiate well between roast_chicken_rice and white_chicken_rice well and between char_kway_teow and mee_goreng.
roast_chicken_rice vs white_chicken_rice
- the issue here is the lack of good images for roast_chicken_rice because the main differentiating feature is the texture of the skin and this would require higher resolution pictures for the CNN to differentiate.
char_kway_teow vs mee_goreng
- the issue here is that both dishes have the thick noodles as its base and both are fried (dry with oily texture). In addition the ingredients are also similar (bean sprouts, vegetables and the occassional egg). I was expecting the CNN to differentiate the colour of the dishes. I believe better higher resolution images would allow the CNN to be better at differentiating between these two
You can download the compiled model here: food-model.hdf5 - Approx 190 MB
Please place the hdf5 model file in the ./dataset folder before running the notebook.
This code is written using TensorFlow 1.3, Keras 2.0.8 and Python 3.6