scarecrow.ai

• 

  Prototype Screenshot

• 

  Healthy vs. Unhealthy

• 

  Applications

• 

  Model Training Part 1

• 

  Model Training Part 2

• 

  Model Training Part 3

• 

  Model Training Part 4

• 

  Model Training Part 5

Team

• Mouhameth T Faye
• Guang Kun Zhang
• Charlotte Maschke
• Sharon Ho
• Ly Bao Phan

Inspiration

To offer a economical solution for smallholder farmers to increase the harvest of healthy crops by automatically identifying plants with diseases & classify them

What it does

Import an image of a plant and use computer vision to determine what type of plant and identify is the plant is healthy or has a type of disease

Challenges

Training the model with a limited amount of data

Testing the model on CPU and it took huge amount of time

Then trying on Nebula AI and use their GPU which is faster

Training the dataset of 44,016 images of plants (leafs) has been used to train the model to recognize what type of plants and what is its conditions:

• 14 different crops species
• 38 class labels
• 26 different diseases

Using fastai.vision to train the model:

• Obtained 99% accuracy on the train model
• Used train model to implement in the test model

Dataset distributions:

• Train: 88%
• Test: 12%

Analysis:

• Learning Rate vs. Loss
• Confusion Matrix
• Creating a mobile friendly interface to show the results
• Getting stuck on how to use the train model to implement on the test model

Built With

• ai
• cnn
• fastai
• flask
• machine-learning
• nebula-ai
• python