Thesis Overview
This thesis investigates how advanced deep learning architectures and ensemble learning can improve the robustness and accuracy of industrial image classification systems. The research was conducted as part of an industry-embedded internship at Nokia Morocco, focusing on AI/ML digital deployment workflows.
Research Questions
- How can advanced CNN architectures be optimized to improve generalization in industrial image classification tasks?
- To what extent does ensemble learning mitigate overfitting and class imbalance in real-world industrial datasets?
Methodology
- Comparative evaluation of CNN architectures (NASNetLarge, InceptionV3, DenseNet, EfficientNet)
- Two dataset structuring strategies to study cross-domain generalization
- Data preprocessing and augmentation to reduce overfitting
- Feature extraction using YOLOv5
- Ensemble learning via prediction aggregation
- Evaluation using accuracy, precision, recall, F1-score, and confusion matrices
Key Results
- Best individual model accuracy: 62.37% (DenseNet169)
- Ensemble model accuracy: 69%, demonstrating improved generalization
- Ensemble learning reduced variance and overfitting compared to single-model approaches
Limitations & Future Work
- Significant class imbalance between “Pass” and “Fail” samples
- Limited dataset size for certain checkpoints
- High computational cost for large CNN architectures
- Future work includes cost-sensitive learning, improved resampling strategies, and hybrid human-in-the-loop feature extraction
Supervision & Affiliation
Academic Supervisor: Prof. Mark Klein (UM6P)
Industry Supervisor: Oussama Ghandari (Nokia)