Residue Segmentation

This project is designed to perform semantic segmentation and classify Residue from Soil for a given set of sugarcane crop images and masks. DeepLabV3plus model is utilized to perform the segmentation.

Results

Original Image - Manual Mask - Predicted Mask

Installation

Use the package manager pip to install XXX.

pip install XXX

Usage

PART 1: Data Setup

The data required for training is obtained using Manually annotated images on LabelBox. LabelBox provides a python API to download images and corresponding masks. Run the following to obtain the images and masks.

python downloadImages.py

This ensures that the data is downloaded and placed in the folders: data/Original_Images and data/Original_Masks

The data needs to be processed into Training - Validation - Test split (Ratio: 80-10-10). Additionally, only the images which are not homogenous will be used in order to enable the model to learn differentiable features. To perform this, run the following.

python ResidueSegmentation_DataPreparation.py

After this, non Homogenous training, validation and testing images will be present at data/Images/Training, data/Masks/Training,data/Images/Validation, data/Masks/Validation,data/Images/Testing, data/Masks/Testing

PART 2: Training

Since the training and validation data is ready, perform training of the model using the following code:

python ResidueSegmentation_Training.py --epochs 5500 --exp_name ResidueSegmentation --model deeplabv3plus --backbone resnet152 --activation sigmoid

Feel free to play around with the input parameters. Eg: model: unet, epochs 11000

This should take a couple of hours depending on the processing capacity of the computer and model used.

Upon completion of training, epoch/100 (default: 55) models + 1 best model (model_best.pth) will be saved in saved_model/exp_name/*

Tensorboard Logs

During training, the IOU score and loss metrics can be visualized in Tensorboard. To view them run the following: (replace exp_name with the appropriate value)

tensorboard --logdir tensorboard_logs/exp_name

The logs would be visible at http://localhost:6006/ on a browser.

PART 3: Testing/Predictions

To perform testing of the test dataset using only the best model, run:

python ResidueSegmentation_Testing.py

This will generate the predicted masks for the test images and store them at outputs/exp_name/*

To Analyse how the model evolves over multiple epochs, GIF images consisting of sequential predicted images are generated for the test data. The logs can also be visualized on tensorboard. To generate this, run:

python ResidueSegmentation_Predictions.py
tensorboard --logdir tensorboard_logs/exp_name

The logs would be visible at http://localhost:6006/ on a browser. The generated GIF image will be present at outputs/GIF/exp_name for each test image.

sandeeppvn/Sugarcane-Semantic-Segmentation-DeepLabV3