ct-hub / pcb-component-detection

Trying to use AI to identify pcb components

PCB-Component-Detection

PCB-CD is one of the engines that in the future will power image2schematic. You can also use it as a standalone tool to identify pcb components. Any help is greatly appreciated!

Testing

• PCB-CD requires numpy, torch, matplotlib and pandas.
• PCB-CD uses a simple label map that can be found in train.py.

training

For training the model, you will need some dataset to work with. The only one I could find is the pcb_wacv_2019 dataset: https://sites.google.com/view/chiawen-kuo/home/pcb-component-detection which includes a number of PCBs labeled with their components. You can use extract_pcb_wacv_2019.py to extract each component to a unique folder. I have uploaded the .csv file but you can also get it with extract_pcb_wacv_2019.py.

extract_pcb_wacv_2019.py assumes you have this folder structure: (there is a one-liner to create all of this below)

├── pcb_wacv_2019 # the downloaded zip file
│   ├── ACM-109_Bottom
│   ├── ACM-109_Top
│   ├── ArduinoMega_Bottom
│   ├── ArduinoMega_Top
│   ...
│   ...
├── pcb_wacv_2019_formatted
│   ├── battery
│   ├── button
│   ├── buzzer
│   ├── capacitor
│   ├── clock
│   ├── connector
│   ├── diode
│   ├── display
│   ├── emi_filter
│   ├── ferrite_bead
│   ├── fuse
│   ├── heatsink
│   ├── ic
│   ├── inductor
│   ├── jumper
│   ├── led
│   ├── potentiometer
│   ├── resistor
│   ├── transformer
│   └── transistor

one-liner:

mkdir pcb_wacv_2019_formatted && cd pcb_wacv_2019_formatted/ && mkdir battery button buzzer capacitor clock connector diode display emi_filter ferrite_bead fuse heatsink ic inductor jumper led potentiometer resistor transformer transistor

Now set imgWriteEnable = True in extract_pcb_wacv_2019.py and run it. You should see all of the components in their respective folder. You could now run train.py, setting training = True. You should now see pcbComponent_net.pth Neural Network model that can be used to predict new samples with predict.py, just set img_path to your image.

Best result so far: (Please zoom in to see detection label)

As you can see, there are a lot of improvement to be done. Any suggestion and help will be greatly appreciated!