Welding Defect Detection with YOLO algorithms on a Limited Dataset: A Comparative Study

Abstract

This work focuses on detecting defects in welding seams using the most advanced YOLO algorithms and transfer learning. To this end, we prepared a small dataset of images using manual welding and compared the performance of the YOLO v5, v6, v7, and v8 methods after two-step training. Key findings reveal that YOLOv7 demonstrates superior performance, suggesting its potential as a valuable tool in automated welding quality control. Our research underscores the importance of model selection. It lays the groundwork for future exploration in larger datasets and varied welding scenarios, potentially contributing to defect detection practices in manufacturing industries. The dataset and the code repository links are also provided to support our findings.

(back to top)

Installation of the environments

For the installation Anaconda is used on Ubuntu 22.04. In command line:

mkdir welding
cd welding

YOLOv5

In the base environment:

conda create --name yolov5 python=3.8 -y
conda activate yolov5
git clone https://github.com/ultralytics/yolov5
cd yolov5
pip install -r requirements.txt

YOLOv7

In the base environment:

conda create --name yolov7 python=3.8 -y
conda activate yolov7
git clone https://github.com/WongKinYiu/yolov7.git
cd yolov7
pip install -r requirements.txt

YOLOv6 and YOLOv8

In the base environment

conda create --name yolov8 python=3.8 -y
conda activate yolov8
git clone https://github.com/ultralytics/ultralytics yolov8
cd yolov8
pip install -e .

Important note: If you have and older GPU, i.e., a GTX 1080 then install an older PyTorch version:

pip install torch==1.12.1+cu113 torchvision==0.13.1+cu113 torchaudio==0.12.1 --extra-index-url https://download.pytorch.org/whl/cu113

(back to top)

Dataset preparations

The full dataset is available at here.

YOLOv5

Download, unzip the welding_images.zip file from the link above which contains the images and the annotations and copy the images directory from it into yolov5/data/ and replace the original images directory. Create a labels directory and copy the .txt files from images to in:

cp images/*.txt labels/.

Copy the yolov5_files/autosplit_*.txt files to the yolov5/data directory.

Copy the yolov5_files/welding_data.yaml file to the yolov5 directory.

Copy the yolov5_files/hyp*.yaml files to yolov5/data/hyps directory.

Example training

python train.py --cos-lr --img 640 --batch 32 --epochs 200 --data welding_data.yaml --weights yolov5n.pt --project defects --name model_5n_dec4 --cache --freeze 10 

YOLOv7

Copy the data directory from yolov5 to yolov7:

cp -avr yolov5/data/images yolov7/data/images
cp yolov5/data/autosplit*.txt yolov7/data/.

Copy the yolov7_files/split_datasplit.py function to yolov7/data and call it by :

python split_dataset.py --folder images --dest images_welding

Copy the yolov7_files/welding_data_v7.yaml file to the yolov7/data directory. Copy the yolov7.yaml and yolov7*.yaml configuration files from the yolov7_files directory to the yolov7/cfg/training directory. Copy the yolov7/utils/loss.py file to yolov7/utils/loss.py. Copy the hyp*.yaml ciles to yolov7/data.

Download from here the yolov7*_training.pt files and copy them to the yolov7 directory.

Copy the yolov7_files/train_aux.py file and copy to yolov7 directory. There has been added a --freeze argument to the parser.

Example training

python train.py --epochs 50 --workers 8 --device 0 --batch-size 4 --data data/welding_data_v7.yaml --img 640 --cfg cfg/training/yolov7_welding.yaml --weights 'yolov7_training.pt' --name yolov7-w1 --hyp data/hyp.scratch.custom.yaml --project defects --freeze 50

YOLOv6 and YOLOv8

We have used the Ultralytics implementation for both models.

Copy the data directory from yolov7 to yolov8 and to yolov6:

cp -avr yolov7/data yolov8/data

Modify the second line if the file /home/$user/.config/Ultralytics/settings.yaml to yolov8 and the erase dataset from the end of line.

Copy the yolov8_files/welding_data_v8.yaml file to the yolov8/data directory. Copy the yolov6_files/welding_data_v6.yaml file to the yolov6/data directory.

Copy the yolov8_files/train.py and yolov8_files/eval.py to the yolov8 directory. (Also, copy the yolov6_files/train.py and yolov6_files/eval.py to the yolov6 directory. These are the same files)

Example training

from ultralytics import YOLO
model = YOLO("yolov8n.pt")
model.train(data="./data/welding_data_v8.yaml",batch=8,imgsz=640,device=0,epochs=100,name="detect/yolov8n_1")

(back to top)

Training and evaluation

The bash scripts for training and evaluation can be found in the script folder. Just copy them to the corresponding yolovx directory.

Example

Example two-phase training for YOLOv5 nano model.

1. Freezing the backbone and train for 200 epochs:

   ./run_training_yolov5n_p5.sh

2. Unfreezeing the backbone and train for 200 epochs:

   ./run_finetuning_yolov5n_p5.sh

3. Evaluating the fine-tuned models on the test set:

   ./run_evaluating_yolov5n_p5.sh

The investigated models and the corresponding filename endings in the scripts directory can be seen in the following table. The scripts are grouped into directories according to the model families, and within that, according to whether we used them for training (train) or evaluation (eval).

(back to top)

Best models

The best finetuned YOLOv7-d6 P6 and YOLOv7x P5 models are available here. These are mentioned in the preprint as the best models in the dataset.

(back to top)