Broccoli head segmentation and classification

This repository containes two different approaches to determine the shape and size of a given broccoli head image given as input:

• Unsupervised Approach
• Supervised Approach

Setup

Create Conda environment

conda env create -f conf/environment.yml

Unsupervised Approach

This approach generates segmentation masks for broccoli heads in an unsupervised fashion. Two different segmentation algorithms are implemented:

• Morphological snakes (ACWE): https://scikit-image.org/docs/dev/api/skimage.segmentation.html#skimage.segmentation.chan_vese
• GrabCut : https://docs.opencv.org/master/d7/d1b/group__imgproc__misc.html#ga909c1dda50efcbeaa3ce126be862b37f

Dataset

The ground truth masks format supported by this implementation is a JSON file containing a polygon of points describing the shape of broccoli heads.

Usage

Run unsupervised/scripts/create_masks.py to create segmented masks given broccoli heads images as input. The script's arguments are:

usage: python unsupervised/scripts/create_masks.py --dataset_path PATH_IMAGES --save_masks_path SAVE_PATH

Run python unsupervised/scripts/evaluate_masks.py to evaluate the predicted segmented masks given broccoli heads images and the ground truth masks as input. The script's arguments are:

usage: python unsupervised/scripts/evaluate_masks.py --dataset_path PATH_IMAGES --number_masks_visualize NUMBER_MASKS

Supervised Approach

Segmentation

This approach generates segmentation masks for broccoli heads in a supervised fashion. A U-net model is impleneted for image segmentation.

Usage

Run supervised/segmentation/scripts/training.py to train a new UNET for broccoli heads segmentation. The script's arguments are:

usage: python supervised/segmentation/scripts/training.py --dataset_images_path PATH_IMAGES --dataset_masks_path PATH_MASKS --save_model_path PATH_SAVE_MODEL

Run supervised/segmentation/scripts/create_masks.py to create segmented masks given broccoli heads images as input. The script's arguments are:

usage: python supervised/segmentation/scripts/create_masks.py --dataset_images_path PATH_IMAGES --model_path PATH_MODEL --save_masks_path SAVE_PATH

Run supervised/segmentation/scripts/evaluate_masks.py to evaluate the predicted segmented masks given broccoli heads images and the ground truth masks as input. The script's arguments are:

usage: python supervised/segmentation/scripts/evaluate.py --dataset_images_path PATH_IMAGES --dataset_masks_path PATH_MASKS --model_path PATH_MODEL --number_masks_visualize NUMBER_MASKS

Binary Classification

This approach includes a binary classifier model, given an image, it detects the presence of a broccoli in the image.

Usage

Run supervised/binary_classification/scripts/training.py to train a new small ResNet for binary broccoli classification. The script's arguments are:

usage: supervised/binary_classification/scripts/training.py --dataset_images_path PATH_IMAGES --save_model_path PATH_SAVE_MODEL

Run supervised/binary_classification/scripts/evaluate.py to evaluate the binary classification model. The script's arguments are:

usage: supervised/binary_classification/scripts/evaluate.py --dataset_images_path PATH_IMAGES --model_path PATH_MODEL

Wrapper Model

This approach implements a wrapper model, which includes a binary classification and segmentation pipeline.

Usage

Run supervised/wrapper/scripts/tile_pipeline.py to segement a tile of broccoli. The script's arguments are:

usage: supervised/wrapper/scripts/tile_pipeline.py --dataset_images_path PATH_IMAGES --annotations_broccoli_path PATH_ANNOTATION --classification_model_path PATH_CLASSIFIER --segmentation_model_path PATH_SEGMENTATION_MODEL --save_masks_path PATH_OUTPUT

Run supervised/wrapper/scripts/evaluate.py to evaluate the predicted segmented masks given broccoli heads images and the ground truth masks as input. The script's arguments are:

usage: supervised/wrapper/scripts/evaluate.py --dataset_images_path PATH_IMAGES --dataset_masks_path PATH_MASKS --classification_model_path PATH_CLASSIFIER --segmentation_model_path PATH_SEGMENTATION_MODEL