Object Detection on component on Handwritten Circuit Diagram, and convert them into LTspice simulation file.

In this notebook, convert a Handwritten Circuit Diagram into LTspice file. using deep learning and image processing, we can extract all the features we need in order to build a simulation file.

the algorithm can be divided into 3 part

1. Deep Learning on a YOLOv4[3] Base model, to extract the components and junctions locations
2. Image processing, to extract all the wire and conation of the components
3. Convert into LTspice file based on all the features we have extracted

In this part, we use a labeled Handwritten Circuit Diagram dataset[1]. we clean it with component we don't have a lot of, and end up with 12 labels.

• capacitor, coil, resistor
• Power and current source
• diode and transistors
• Intersections and overleap on wires (Two different types of each)
• and finaly, text

We Fine-Tuning a YOLOv4[2] base model, that was train on general object detection. and got good results with a good yolo loss:

we used in order to train the model:

• Epochs: 70
• Batch size: 2
• Subdivisions: 1
• Learning rate: 0.0001
• Optimizer: adam
• Loss function: yolo_loss
• Dataset classes: 12

to extract the wire and conetion between the components. we use a basic image processing techniques, including:

1. convert the image into binary image
2. Erode and Dilate to remove noise
3. Remove label components from the image
4. Find Contours
5. Find connection using the contours

after we remove the labeled components from the images, we should stay only with the wire that conect the components. we can easy find them with "Find Contours", and use them to conect 2 compunent with the same contour.

finally with all the Features, we can code circuit using the location, and the connection of the components.

We train the model on google colab, and so the can be easly rerun. how to run it is really written in the ipynb file. see the Test file to test the project.

• training file - TRAIN-Circet_Object_Detector_YOLOv4.ipynb
• testing file, with the best weights - TEST_with_image-Circet_Object_Detector_YOLOv4.ipynb
• code to clean the dataset - basic_work_with_dataset.py
• the model - models.py
• the code to train the model - train.py

• the code that create component in an LTspice file - component_pleaser.py
• image proccesing code, that also create the LTspice file - build_circet.py (get as an input the labels of the image from the model, and the image)

[1] Hand drawing circet diagram dataset - https://arxiv.org/abs/2107.10373

[2] the model we toke and fine tuning - https://github.com/roboflow-ai/pytorch-YOLOv4

[3] Yolov4 model article- https://towardsdatascience.com/yolo-v4-optimal-speed-accuracy-for-object-detection-79896ed47b50#6ac9