Defect-Detection-yolov7

This project is based on yolov7 project. The original project can be found here

This model are designed for a subproject of defect detection. It can detect Multiple equipment components in single image or video stream.

Requirements

• Python 3.8 or later with all [requirements.txt]
• PyTorch 1.7.0 or later with CUDA and torchvision

create virtual environment

python -m venv Defect-Detection
source Defect-Detection/bin/activate 

install requirements

cd $project_path
pip install -r requirements.txt

Raw data processing

Data should store in yolov7-format dir tree as follow:

yolov7_data
└── sub_yolo_data
  ├── test
  │   ├── images
  │   └── labels
  └── train
      ├── images
      └── labels

Where images should be .jpg format and labels should be yolo .txt format.

For raw data in 缺陷识别 folder, way to process data is as follow:

Seperate mixed data

In each data folder,images(.jpg files and .JPG files) and labels(.txt files) are mixed together. We need to seperate them into two folders and rename them.

Code for replacing space with underscore

data_dir="~/Documents/Defect-Detection-yolov7/data/缺陷识别raw/设备部件识别"
cd $data_dir
for f in *; do mv "$f" `echo $f | tr ' ' '_'`; done

Code for seperating images and labels

data_dir="~/Documents/Defect-Detection-yolov7/data/缺陷识别raw/设备部件识别"
touch mvjpg.txt
touch mvxml.txt
find $data_dir -iregex .*\.jpg    > mvjpg.txt
find $data_dir -iregex .*\.xml    > mvxml.txt
mkdir $data_dir/images
mkdir $data_dir/labels
for i in $(cat mvjpg.txt); do (mv $i $data_dir/images); done
for i in $(cat mvxml.txt); do (mv $i $data_dir/labels); done
rm mvjpg.txt
rm mvxml.txt

Deal with images with wrong label in the dataset

Labeling errors are common in the dataset, so we need to deal with them. For some unknown reason, labels of some images are rotated by 90 degrees(seems randomly clockwise or counterclockwise).

# find the images with wrong label
python data/wrong_size.py

The outout is a txt file with the path of images with wrong label.including rotated images and images with incorrect label for other reasons.then seperate them from the dataset.

data_dir="/Users/wzilai/Documents/Defect-Detection-yolov7/data/缺陷识别raw/设备部件识别"
mkdir $data_dir/bad_examples/images
mkdir $data_dir/bad_examples/labels
for i in $(cat $data_dir/wrong_size.txt); do mv $data_dir/images/$i $data_dir/bad_examples/images; done
for i in $(cat $data_dir/wrong_size.txt); do mv $data_dir/labels/$i $data_dir/bad_examples/labels; done

You can fix the label of these images manully or just use the remaining images to train the model.

Change label format

raw labels are stored in xml format,but yolo model need input in yolo-format. prepare.py can change xml format to yolo-format and separate data into train set and validation set. before run prepare.py, you need to change class name and yolo-format dir name in prepare.py. dir create is not fully-automatic in prepare.py, you need to create dir by yourself for now.

data dir tree

yolov7_data
└── sub_yolo_data
  ├── test
  │   ├── images
  │   └── labels
  └── train
      ├── images
      └── labels

Create yaml file

yaml file is used to record the project information. After run prepare.py, create yaml file for dataset, then change train/test image path and class name in yaml file. train.py will automatically read yaml file, and find labels.

Sample yolov7 yaml file

train: ./data/dataset/train.txt or ./data/dataset/train
val: ./data/dataset/val.txt or ./data/dataset/val
test: ./data/dataset/test.txt or ./data/dataset/test
#txt files were auto generated by prepare.py
# Number of classes
nc: 3
# Class names
names: ['cat', 'dog', 'person']

training

track and visualize

yolov7

!pip install wandb
import wandb
wandb.login()

If you get fresh copy from yolov7 github, please modifie Defect-Detection-yolov7/utils/google_utils.py delete dtype=np.int32 in cum_counts = np.cumsum(np.greater(counts, 0, dtype=np.int32)) or change int32 to float32 because this typecheck will cause error in wandb

train

!python train.py --epochs 300 --data 设备部件识别_without_ddjt_yx_tg.yaml --weight yolov7-tiny.pt --batch-size 16 --freeze 50 --hyp data/hyp.scratch.tiny.yaml

you can change hyperparameters in hyp.yaml file.

result for 300 epochs:

6085                  all         215         221       0.478       0.548       0.436       0.211
6086             cysb_sgz         215           2       0.129         0.5       0.108      0.0557
6087               SF6ylb         215           1       0.512           1       0.995       0.498
6088                  drq         215          55       0.433       0.327       0.194      0.0742
6089                ecjxh         215           5       0.408         0.8       0.524       0.241
6090                drqgd         215           7       0.207       0.286       0.116      0.0659
6091             cysb_lqq         215          21       0.453       0.619        0.48         0.2
6092           cysb_qtjdq         215           6       0.437           1       0.828       0.371
6093                xldlb         215          22        0.66           1        0.98       0.477
6094                  ywj         215           1       0.204           1       0.498       0.299
6095                jdyxx         215          13       0.481       0.154       0.116      0.0195
6096                fhz_f         215           1           1           0           0           0
6097                 bmwh         215           3           0           0       0.112      0.0335
6098               xmbhzc         215           3           1       0.664       0.736       0.442
6099                  pzq         215          19       0.478       0.842       0.687       0.308
6100                  jyh         215          20       0.478         0.8       0.705       0.366
6101                  ywc         215          11       0.118       0.182       0.144      0.0641
6102             cysb_cyg         215          30       0.598       0.695       0.634       0.292
6103                 bjzc         215           1           1           0           0           0

Detect

!python detect.py  --weights runs/train/exp10/weights/best.pt --source data/缺陷识别yolov7/设备部件识别_without_ddjt_yx_tg/test/images

Usage - sources:
    --source 0                               # webcam
            img.jpg                         # image
            vid.mp4                         # video
            screen                          # screenshot
            path/                           # directory
            list.txt                        # list of images
            list.streams                    # list of streams
            'path/*.jpg'                    # glob
            'https://youtu.be/Zgi9g1ksQHc'  # YouTube
            'rtsp://example.com/media.mp4'  # RTSP, RTMP, HTTP stream

Example