search

dahalsweekar / Deep-Weed-Segmentation

This repository offers an implementation of diverse segmentation models for semantic segmentation. The provided method allows the integration of different networks and backbones to create a combination of choices.
MIT License
0 stars 1 forks source link

readme

Deep-Weed-Segmentation

Introduction

This repository offers an implementation of diverse segmentation models designed for classifying weeds into four distinct categories. The provided method allows the integration of different networks and backbones to create a combination of choices.

For more detail, please refer to our paper: Deep Learning-Based Weed Detection Using UAV Images: A Comparative Study

Task List

Features

Training/Evaluation

Flags Usage
--network Define network (Default: custom)
--backbone Define backbone (Default: None)
--patch_size Define patch size (Default:256)
--weight_path Set path to model weights
--data_path Set path to data
--epoch Set number of epochs (Default: 50)
--verbose Set verbose (Default: 1)
--batch_size Set Batch size (Default: 8)
--validation_size Set Validation size (Default: 0.1)
-test_split Set test size (Default: 0.2)
--visualizer Enable visualizer (Default: Not enabled)
--score Enable score calculation after training (Default: Not enabled)
--test Enable testing after training (Default: Not enabled)
--binary Enable class 2 training (Default: Not enabled)
--augment Enable Augmentation (Default: Not enabled) WARNING! May cause system to crash
--threshold Set threshold value (Default: 0.03)
Network BackBone
custom |None
unet | vgg16, resnet50, inceptionv3, efficientnetb0, densenet121, mobilenetv2
linknet | vgg16, resnet50, inceptionv3, efficientnetb0, densenet121, mobilenetv2
pspnet | vgg16, resnet50, inceptionv3, efficientnetb0, densenet121, mobilenetv2
segnet | vgg16, resnet50, inceptionv3, efficientnetb0, densenet121, mobilenetv2
deeplabv3 | vgg16, resnet50, inceptionv3, efficientnetb0, densenet121, mobilenetv2

for pspnet image size must be divisible by 48, the image size will be adjusted accordingly.

Installation

Requirements

-Python3
-Cuda

Install

1. git clone https://github.com/dahalsweekar/Deep-Weed-Segmentation.git
2. pip install -r requirements.txt

Training

Training is set to early stopping

python services/train.py --network unet --backbone vgg16 --patch_size 128 --batch_size 4 --epoch 20 --score --data_path /content/drive/MyDrive/data/CoFly-WeedDB

Models

Trained models are saved in ./models/

Dataset

Root of the dataset, by default, is ./data/CoFlyWeed-DB/

|
|
|__./data/CoFlyWeed-DB/
|
|___/images
|
|__*.jpg .png*
|
|___/labels_1d
|
|__*.jpg .png*

Evaluation

A model must be trained and saved in ./models/ folder first

python services/eval.py --network unet --backbone vgg16

Third-Party Implementations