Skip to content

Latest commit

 

History

History
174 lines (134 loc) · 7.49 KB

README.md

File metadata and controls

174 lines (134 loc) · 7.49 KB
Raw

Weeds Detection Using Tensorflow

Table of Content

Introduction

This software is build using tensorflow framework to detect the most common weeds in the rural environment. Due to overhead of GPU memory (6G RTX 2060) on my local desktop, I only trained faster_rcnn_inception_v2_coco from tensorflow object detection models with three type of weeds (dandelion, oxalis,buckhorn plantain). you can definitely train the model to detect more weeds on your computer with much better GPU. The google search engine also built in software allow users to check out solution to kill that specified weed in real time. Here is a quick demo:

demo video

Environment Requirements:

  • Linux (Ubuntu)
  • cuda >=9.0
  • Python >= 3.5
  • Tensorflow-gpu>=1.4.0
  • PyQt5
  • opencv-python3
  • pillow
  • jupyter notebook
  • matplotlib

Prepare and Label image data:

Download images

All weed images are located in corresponding directory within images folder. In case that you want to try out different type of weeds, you can download more images via google image download (make sure that images are placed in directory under their name within images folder !).

Label image data

Download labelImg from (tzutalin/labelImg), then follow through all procedures and install on your computer. Once all installations are finished, go to labelImg folder and type this command in your terminal:

$python3 labelImg.py

After entering above command, A GUI window will pop up like this:

alt text

Click the Open Dir and load all images into window. Once this is done, the bottom right corner will list all uploaded images
alt text

Click Create\nRectBox (Keep PascalVOC format !!!) and start to label image.

demo video

After the image is labeled, and press Ctr+S to save all labeled data into xml file. after it is saved, Click Next images and repeat same process

Generate tfrecord:

Split all labeled images along with their xml files(80% for training and 20% for testing) and place them into test and train in images folder. And convert all *.xml to csv files by running python scripts from your terminal:

$python3 convert_xml_to_csv.py  --test_input images/test  --train_input  images/train

Now we need to create weed label map in data folder:

item{

    id:1
    name: 'buckhorn plantain'
    display_name: 'buckhorn plantain'
}

item{

    id:2
    name: 'dandelion'
    display_name: 'dandelion'
}

item{

    id:3
    name: 'oxalis'
    display_name: 'oxalis'
}

Before generating tfrecord file, place following codes in the function called class_text_to_int in generate_tfrecord.py (same as label map we just created)

def class_text_to_int(row_label):
    
    if row_label == 'buckhorn plantain':
        return 1
    elif row_label == 'dandelion':
        return 2
    elif row_label == 'oxalis':
        return 3
    else:
        None

Next, generate tf record file for test and train data just by run python script from your terminal:

$python3 generate_tfrecord.py\
        --csv_input=data/test_labels.csv\
        --output_path=data/test.record\
        --image_dir=images/test/

$python3 generate_tfrecord.py\ 
         --csv_input=data/train_labels.csv\
         --output_path=data/train.record\ 
         --image_dir=images/train/

After running above scripts, the tfrecod file for both train and test should be located in data folder.

Training The Model:

To train tensorflow model, git clone https://github.com/HarryChen1995/tensorflow_model_API or download the .zip and extract it to your desktop. Then go to research folder and compile object_detection with protobuf by enter following command from terminal:

$ protoc object_detection/protos/*.proto --python_out=.

then export slim python path

$ export PYTHONPATH=$PYTHONPATH:`pwd`:`pwd`/slim

Note: this command need to run for every time you start new terminal. To avoid running manually, placing following bash script in your ~/.bashrc:

export PYTHONPATH=$PYTHONPATH:/home/username/Desktop/tensorflow_model_API/research:/home/username/Desktop/tensorflow_model_API/research/slim

and compile ~/.bashrc from terminal:

$ source ~/.bashrc

Finally install tensorflow model API by enter following command in research folder:

$python3 setup.py install

Now we need to download physical model (faster_rcnn_inception_v2_coco) from Tensorflow detection model zoo and extract it to object detection folder.

Next, copy data and training folder to object detection folder, then proceed to train model by entering following command from terminal within object detection folder:

$python3 train.py --logtostderr --train_dir=model_output/ --pipeline_config_path=training/faster_rcnn_inception_v2_coco.config

Note: it may take up to 30-60 seconds to initialize tensorflow, Once initialization finished, the training output for each step will be displayed in terminal: alt text

Depend on GPU and CPU of your computer, the whole training process may take up to couple hours. I recommend allowing to train your model until the loss get down below 0.05. After training is finished, you can check total loss of model through tensorboard by running following command within object detection folder:

$ tensorboard --logdir=model_output

Then open your browser and go to localhost:6006, the loss will be displayed along with its each training step in browser:

alt text

Export Frozen Inference Graph:

Once the training is completed, then generate inference model by enter following command from terminal within object detection folder:

    python3 export_inference_graph.py \
    --input_type image_tensor \
    --pipeline_config_path training/faster_rcnn_inception_v2_coco.config \
    --trained_checkpoint_prefix /model_output/model.ckpt-latest_checkpoints \
    --output_directory weed

After this, the last step is to move weed folder into weed_detections folder (we are done finally !!!!)

Usage:

To run software and enter following command from terminal within weed_detections folder

$python3 user_gui.py