Computer Vision for Tea Leaves

Analysis 

We performed experiments using VGG and Resnet18 neural networks with transfer learning(TL) and data augmentation(DA) on RGB and gray scale images with a split of 70% training, 15% validation, and 15% test. We used early stopping in all the experiments.

Brainstorming

• Led a team of 7 where we exchanged ideas

• Each member shared an idea of a dataset and model

• We compared options

• Selected the most voted project

Computer Vision Model

• Transfer learning

• Data Augmentation

• Modeling and experimentation

• Determined the variables and hyperparameters

• Planning/drafting the model

• Implementing the model

• Testing and evaluating the model

• Publishing the model

Results

From some global metrics, we found out that the highest recall value was 0.92 for the resnet18(RGB) + DA + TL when using the test dataset.

Also, we discovered that the Alga leaf disease class showed the highest accuracy among the 7 classes for the VGG RGB network whereas the gray light class showed the highest accuracy for the Resnet18(RGB) + DA + TL.

Lastly, we found that 2 classes had the highest accuracy in both models.

Conclusions 

In conclusion, the best results were obtained using restnet18 RGB with data augmentation and transfer learning. The hardest challenge I had was to persuade the team of 4 engineers and scientists to get to the goal in such a short period of time.