FoodVision

Deep learning experiments for the image classification starting from 10 classes to 101 food classes.

Overview

For this project I built a food classifier to identify food from different food images. This project aim was to beat 77.4% top-1 accuracy of DeepFood paper. I was able to get the model to predict the class of the food from 101 classes with 70% accuracy with out fine-tuning and 80% with fine-tuned model. To get these results I used transfer learning (no fine-tuning) on a existing noncomplex pretrained (on ImageNet) CNN model "EfficienetNetB0". This created time efficiencies and solid results. Further experiments are needed to improve the accuracy like using more complex pretrained models (EfficienetNetB4 or ResNet34).

Code and Resources Used

• Python Version: 3.8
• Tensorflow Version: 2.7.0
• Packages: pandas, numpy, sklearn, matplotlib, seaborn, selenium, flask, json, pickle
• GPU Configuration: Using Google Colab

Dataset and EDA

The Food101 datasets was downloaded from Kaggle.

	Exp #1: 10-Classes	Exp #2: 101-Classes
Dataset source	Preprocessed download from Kaggle	TensorFlow Datasets
Train data	7,575 images	75,750 images
Test data	25,250 images	25,250 images
Data loading	tf pre-built function	tf.data API
Target results	50.76% top-1 accuracy (beat Food101 paper)	77.4% top-1 accuracy (beat DeepFood paper)

some examples from the 101Food datasets:

Model Building and Performance

	Exp #1: 10-Classes	Exp #2: 101-Classes
Models	EfficienetNetB0	EfficienetNetB0
Pretrained	on ImageNet	in ImageNet
Fine-Tuned	Yes	Yes
Fine-Tuned layers	last 5	all
Accuracy	57.84%	80.%
Training Time	~5m	~20m

Confusion Matrix (101-Food Classes, Exp #1)

F1-score across classes:

Top few wrong predictions with high prediction probs:

It is clearly visible that most wrong prediction with high pred probs are not generalizable with human brain also.

Predictions on Custom Images

The model predicts all the 10 images correct with high probability score. These images were collected from different sources.