This repository contains the code for the solution I submitted for the Kaggle Plant Pathology 2021 challenge, which took place from March 15 2021 to May 27 2021. This competition was part of the Fine-Grained Visual Categorization FGVC8 workshop at the Computer Vision and Pattern Recognition Conference CVPR 2021.
A corresponding article can be found on my website here.
This competition was a good opportunity to explore some technical topics related to Convolutional Neural Networks and computer vision such as :
- How to implement a CNN taking advantage of TPUs to speed up the computing steps ;
- How to build an efficient TensorFlow input pipeline with the
tf.dataAPI ; - What loss could be suitable for optimizing the F1-score ;
- What are Vision Transformer neural networks.
My solution ranked 11th out of 626 teams on the public leaderboard, and 36th on the private leaderboard (top 6%).
As stated on the competition description page :
"Apples are one of the most important temperate fruit crops in the world. Foliar (leaf) diseases pose a major threat to the overall productivity and quality of apple orchards. The current process for disease diagnosis in apple orchards is based on manual scouting by humans, which is time-consuming and expensive."
The task of this challenge was thus to develop a machine learning-based model to identify diseases on images of apple tree leaves.
Each leaf could be healthy, or present a combination of various diseases. As each image could potentially be associated with several labels (in case of multiple diseases), this was a multi-label classification task.
For the purpose of the competition, a dataset of 18632 labeled apple tree leaf images was provided.
The test set used to evaluate the participant submissions was constituted of roughly 2700 images.
The pictures were provided in jpeg format of relatively high resolution, lots of them being 2676 x 4000 pixels, but the resolution and aspect ratio could somewhat vary for some images.
The evaluation metric for this competition was the Mean F1-score.
My best score was reached by averaging the output of 3 different models :
- a ResNet50,
- an EfficientNetB7,
- and a Vision Transformer model.
On top of the training process optimizations, significant results improvements were brought by :
- Suitable image augmentation,
- Handling the cases were no label has been predicted by the model (probability of every label inferior to the chosen threshold),
- Test Time Augmentation (TTA) (see this article for a brief explanation on how it works).
4 Notebooks are available in this repository :
- 1_plant-pathology-2021-fgvc8-resnet50-training.ipynb explains the training of the ResNet50 model in details ;
- 2_plant-pathology-2021-fgvc8-efficientnetB7-training.ipynb contains the code of the EfficientNetB7 model training with only brief explanations as the steps are much the same as for the ResNet50 model ;
- 3_plant-pathology-2021-fgvc8-vit-training.ipynb contains the code of the Vision Transformer model training with only brief explainations as the steps are much the same as for the previous two model ;
- 4_plant-pathology-2021-fgvc8-inference.ipynb contains the code of the inference process.