This project code is forked from https://github.com/DetectionTeamUCAS/FPN_Tensorflow. I have only made minor changes on this wonderful and clear project. Thanks for their perfect code. I can learn and apply it to a new problem.
The Open Lab on Human Robot Interaction of Peking University has released the PCB defect dataset. If you need to do experiments on this dataset, please contact the dataset author.
6 types of defects are made by photoshop, a graphics editor published by Adobe Systems. The defects defined in the dataset are: missing hole, mouse bite, open circuit, short, spur, and spurious copper. For example:
The paper of this project will be update. However, the defect images of raw dataset are high-resolution. With the respect of such small dataset, data augmentation techniques are adopted before data training. The images are then cropped into 600 × 600 sub-images, forming our training set and testing set with 9920 and 2508 images, respectively.
You can download augmented dataset:
(1) For Baiduyun: https://pan.baidu.com/s/1eAxDF4txpgMInxbmNDX0Zw code: a6rh
(2) For Dropbox: https://www.dropbox.com/s/h0f39nyotddibsb/VOC_PCB.zip?dl=0
The augmented dataset contains 10668 images and the corresponding annotation files.
1、This augmented dataset is privately owned, if you want to use it in your paper, please contact me.
2、If you have questions, please open an issue, I will reply as soon as possible. If you don't have some difficult problem about this project, maybe you don't need to send me an email and add wechat, just open an issue.
3、Of course, if you have other research questions like dataset and defect detection, welcome to send me an email.
4、Cite: Ding, Runwei, et al. "TDD-Net: A Tiny Defect Detection Network for Printed Circuit Boards." CAAI Transactions on Intelligence Technology (2019).
cd $PATH_ROOT/libs/box_utils/cython_utils python setup.py build_ext --inplace
Then download the corresponding weights:
For Baiduyun: weights. code：shac
For Google driver: weights
And then put the trained weights into
TRAINED_CKPT = os.path.join(ROOT_PATH, 'output/trained_weights')
It depends on your project path. Please read '$PATH_ROOT/libs/configs/cfgs.py'.
cd $PATH_ROOT/tools python inference.py --data_dir='/PATH/TO/THE/TO/BE/DETECTED/IMAGES/' --save_dir='/PATH/TO/SAVE/RESULTS/' --GPU='0'
After running this code, you will get the detected image in your 'save_dir' path.
If you want to train your own data, please follow this project: https://github.com/DetectionTeamUCAS/FPN_Tensorflow
1、Select a configuration file in the folder ($PATH_ROOT/libs/configs/) and copy its contents into cfgs.py. Modify parameters (such as CLASS_NUM, DATASET_NAME, VERSION, etc.) in $PATH_ROOT/libs/configs/cfgs.py
2、Generate train and test floder
cd $PATH_ROOT/data/io/ python divide_data.py
You should check the image_path and xml_path in the 'divide_data.py'
(1)Modify parameters (such as VOC_dir, xml_dir, image_dir, dataset, etc.) in $PATH_ROOT/data/io/convert_data_to_tfrecord.py
cd $PATH_ROOT/data/io/ python convert_data_to_tfrecord.py
cd $PATH_ROOT/tools python train.py
cd $PATH_ROOT/tools python eval.py --eval_imgs='/PATH/TO/THE/TO/BE/EVALED/IMAGES/' --annotation_dir='/PATH/TO/TEST/ANNOTATION/' --GPU='0'
After running this code, you will get the precision, recall and AP of per defect type.