Situation
Manual visual inspection of aluminium castings is inconsistent and slow, and manufacturing teams need a fast first-pass check that can flag likely defects for review.
Data Science
Casting Defect Detection
A CNN image-classification workflow for identifying defective industrial castings from visual inspection data.
View source repository↗
Why it matters
Shows how deep-learning classification can support faster and more consistent manufacturing quality checks.
My role
Computer-vision practitioner
- 01Built a consistent grayscale image pipeline with 128×128 resizing and pixel normalization for training and single-image inference.
- 02Developed a compact regularized CNN with augmentation, convolution/pooling blocks, dropout, sigmoid output, and early stopping.
- 03Evaluated the classifier on validation and held-out test data and packaged reusable prediction and model-saving workflows.
Task
Build a binary image classifier that distinguishes defective casting fronts from acceptable parts, with a preprocessing and inference pipeline consistent enough to trust on new images.
Action
- Images from def_front and ok_front folders are converted to one-channel grayscale, resized to 128×128, normalized to the [0,1] range, and split into train, validation, and test sets.
- A compact Keras CNN applies optional random flips and rotations, repeated Conv2D/MaxPooling2D blocks, dropout regularization, and a sigmoid output, trained with Adam and binary cross-entropy, with early stopping on validation loss.
- Training curves, held-out test loss, a confusion matrix, and a full classification report assess overall performance and class balance rather than a single accuracy number.
- A single-image inference helper applies the identical grayscale, resize, and normalization pipeline and returns the predicted class and defect probability, with the trained model saved for reuse without retraining.
Result
The final model reached approximately 95% validation accuracy and 96% test accuracy, with precision and recall between 0.93 and 0.96 across both defective and acceptable classes.
Technical implementation
How the solution was built.
Each layer connects an implementation choice to the decision or workflow it supports.
06 layers| Layer | Implementation | Operational purpose |
|---|---|---|
| Image pipeline | Grayscale conversion, 128 by 128 resizing and pixel normalization | Apply identical preprocessing during training and inference |
| Data split | Separate training, validation and held-out test datasets | Tune the workflow without contaminating final evaluation evidence |
| Augmentation | Controlled random flips and rotations | Improve robustness to small visual variations in part orientation |
| CNN architecture | Conv2D, MaxPooling2D, dense layers, dropout and sigmoid binary output | Learn local defect patterns with a compact laptop-trainable model |
| Training control | Adam, binary cross-entropy and early stopping on validation loss | Optimize classification while limiting overfitting |
| Evaluation and reuse | Confusion matrix, classification report, single-image prediction and saved model artifacts | Verify class-level performance and support repeatable inference |