This article was originally written for Meritocracy.is
Food security is one of the United Nations Sustainable Goals. Yet still, every day more than 25,000 people die of hunger. Overall, there are more than 800 million undernourished people in the world because of food shortages or inefficient crop yield. Early disease detection on plant leaves can reduce world's crop losses.
In this project I looked at beans dataset, where each image contains bean plants grown on small farms in East Africa. Beans are a staple in many East African cousines and represent a significant source of protein for school-aged children. Each image from the beans dataset is associated with exactly one condition:
- angular leaf spot disease
- bean rust disease
- healthy
Our goal is to develop a classifier that can predict one of these conditions. Every image in this dataset is 500-by-500 pixels large and was taken by a smartphone camera on the farm.
This code contains a basic convolutional neural network architecture trained from scratch and achieves 75 % accuracy on the test set (so far).
- Network depth: what happens if we add or remove convolutional layers?
- Image augmentation: what if we apply geometric transformations, such as scaling, cropping, to training images to add diversity and increase the training set size?
- Image resizing: what is we crop images with
tf.image.resize_with_crop_or_padto reduce background noise? - Fine-tuning the hyper-parameters: what if we try other hyperparameter values?