This GitHub repo presents a Deep Learning model for PCB image classification, achieving high accuracy. We provide a comprehensive dataset and explainability analysis. Our work contributes to the field of image classification and explainable AI. Code and resources are available for use and further research.
In particular the repo contains:
-
model: a folder containing everything used to train the model (ResNet50d):
- main: the train
- data: data augmentation tranformations, functions used to load dataset into dataloaders (train-test-validation)
- optimizer: RAdam code
- utils: some useful functions
-
notebooks: a folder containing the notebooks used to check and plot results:
- plot: results of the training
- analysis: tests of the model
- masking_images: this notebook contains the passages applied to masking out relevant areas. First of all the areas corresponding to the relevant parts highlighted by CAM-based techniques. Then the areas corresponding to the failed components of the pictures segmented using external tools.
-
cam: a folder containing the notebook used to apply CAM-based techniques.
-
examples: a folder containing some examples of the steps applied:
- defective folder: contains ten randomly selected pictures of "defective" PCB, which are correctly classified by the model
- cam_results: contains an example of a portion of board correctly defined as "defective" analyzed with the CAM-based techniques
- masked_cam: contains an example of a portion of board where areas detected by the CAM-based techniques are masked out
- masked_components: contains an example of a portion of board where failed components are masekd out
- gradCAM: contains the analysis of ten randomly selected pictures classified as "defective" by the model analyzed using GradCAM