Hybrid Defect Detection for Ceramic Tiles
Tiny-defect detection on ultra-high-resolution industrial images with an improved YOLOv8.
Group Leader — Course Project, JHU 601.661 Computer Vision · Sep 2025 – Dec 2025
- Sliced 8192×6000 industrial images via sliding windows into a ~19.5K-image localized dataset to mitigate small-object feature loss.
- Improved YOLOv8n with CBAM attention, SPD convolution, a custom Weighted CIoU-NWD loss, and a C2f-Faster (Partial Convolution) backbone.
- Achieved +5.2% mAP → 71.2% while maintaining 133 FPS at only 3.10M parameters.