gear-teeth-defect-detection-YOLOv3

Download YOLOv3 weights at the very first step

$ wget https://pjreddie.com/media/files/yolov3.weights

---

Annotation the images with LabelImg

• To Install LabelImg -

git clone https://github.com/heartexlabs/labelImg.git

conda create -n LabelImg python=3.9.13

conda activate LabelImg
pip install labelimg

• Open LabelImg

# activate the conda environment
conda activate LabelImg

# cd to the cloned labelImg-Repo
ls
cd lebelImg-Repo
ls

# open labelImg tool
python labelImg.py

• Annotation procedure;

---

convert annotated xml to csv_file

# For training-set
python xml_to_csv.py --path_to_xml /path/to/train_images_folder --path_to_csv /path/to/train_images_folder/annotation.csv

# For testing-set
python xml_to_csv.py --path_to_xml /path/to/test_images_folder --path_to_csv /path/to/test_images_folder/annotation.csv

Create labelmap.pbtxt

/img_xml_data/labelmap.pbtxt

---

Convert csv_file to annotation.txt file in a separate data folder(yolov3_data)

# For train_gear_annotations.txt
python prepare_data.py --path_to_images /path/to/train_images_folder --path_to_csv_annotations /path/to/train_images_folder/annotation.csv --path_to_save_output /yolov3_data/train

# For test_gear_annotation.txt
python prepare_data.py --path_to_images /path/to/test_images_folder --path_to_csv_annotations /path/to/test_images_folder/annotation.csv --path_to_save_output /yolov3_data/test

---

Adding class names

/classes/gear_teeth.names

Changing the necessary parameters in configuration file

/core/config.py Change the preparameter & hyperparameters for model training based on the machine being trained on;

• add the class name path ('/classes/gear_teeth.names') / __C.YOLO.CLASSES
• training annotation path ('/dataset/train_gear_annotations.txt') / __C.TRAIN.ANNOT_PATH
• training batch size (depending on GPU size) / __C.TRAIN.BATCH_SIZE
• training input size of neurons (depending on GPU size) / __C.TRAIN.INPUT_SIZE
• data augmentation (True or False) / __C.TRAIN.DATA_AUG
• initial and final learning rate / __C.TRAIN.LR_INIT, __C.TRAIN.LR_END
• numbers of epochs / __C.TRAIN.EPOCHS
• testing annotation path ('/dataset/test_gear_annotations.txt') / __C.TEST.ANNOT_PATH

Note Model accuracy & performance will be depending on some of the hyperparameters such as epochs, batch size, neuron sizes and learning rate.

---

Training YOLOv3 Object Detection Model

After the changing and adding some parameter in configuration file.

python train.py

---

Analyzing the results of training

After the model had been trained, the performance of the model was visualized and analyzed on Tensorboard using the trained log.

tensorboard --logdir './result_output/log'

---

Evaluating the trained model on the testing set

Test the trained model on the testing set of data, and then check the result of tested image with bounding boxes in ./result_output/eval_detection.

python test.py

---

Compute mean average precision (mAP) of the trained model

To compute mAP of the trained model;

python mAP/main.py

The computed mAP can be checked in ./results/mAP.png.

The mAP on each class achieved the best result with accurately detected score.

---

Save the model

The trained weights of the model is saved as TF2 model format for production ready by running

python ckpt_to_savedModel.py

After running, the trained & production-ready model will be save in ./SavedModel/YOLOv3_model/

---

Inferencing the model on inferencing data (new images)

To inference the model;

python run_inference.py --path_to_images './inference_data/'

Sample images after inferencing

---

References;

https://github.com/sniper0110/YOLOv3

https://github.com/YunYang1994/TensorFlow2.0-Examples/tree/master/4-Object_Detection/YOLOV3