The project is designed to detect and segment defects in PCBs (Printed Circuit Boards) utilizing the SegFormer architecture, aiming to enhance the quality standards in manufacturing processes. Project covers a series of steps such as data visualization, data preprocessing, model training, and inference.
The project requires all the packages requirements.txt specified.
The dataset provided in the data folder, consists of images and their corresponding annotations in XML format. It provides mid & high-resolution images of PCBs and each of the annotated to indicate defects.
For an in-depth understanding of the dataset & structures and to utilize it in different formats visit Roboflow.
- Set up the environment and install the necessary packages.
- Import the required libraries.
- Define the data transformations and create a custom dataset class for loading the data.
- Set up the training arguments and define the metrics for evaluation.
- Train the Segformer model using the Hugging Face's Trainer class.
- Perform inference on the validation set and visualize the segmentation results.
The notebook includes, Visualizations: Defect Heatmap, Defect Size Distribution, examples of Segmented Outputs and the Model Outcomes.
The model showcased in this project was trained on a relatively small dataset but still managed to achieve solid results. If you aim to further enhance the model's performance, consider:
- Augmenting the Data: Implement advanced data augmentation techniques such as rotations, zooming, and color jittering to artificially increase the dataset's size.
- Fine-tuning: Use a pre-trained model and fine-tune it on your specific dataset for better generalization.
- Regularization: Implement techniques like dropout or weight decay to prevent potential overfitting due to the smaller dataset size.