Abstract:
Ceramic tiles are a cornerstone in construction and interior design, thereby indicating that the quality of surface affects directly its durability and appearance potential. Detection of surface defects on tiles, such as cracks, holes or edge chipping are commonly performed by relying almost exclusively on human operators. These labor-intensive inspections are both inconsistent and susceptible to error. To cope with this issue, in the present work we propose a deep learning-based defect detection using a CTDD-YOLO architecture, an upgraded network of YOLO family specifically for lightweight and fast object detection. The model was trained and tested on a Roboflow dataset publicly accessible, including images of ceramic tiles with three major types of defects: edge chipping, holes and line scratches. Model training and finetuning We trained and finetuned our model using Ultralytics YOLOv8 framework on GPU resources in Google Colab. Evaluation was done using common object detection metrics such as Precision, Recall and mAP values. The learned system showed that it can effectively detect incidents with a Precision of 66.7%, Recall of 61.9% and mAP@50 = 64.8%, indicating an acceptable detection performance which is also computationally efficient. The experimental results demonstrate that CTDD-YOLO can effectively realize the automation of tile inspection without relying on the manual identification in industrial production lines. This work not only verifies the application potential of deep learning to surface defect detection but also paves a way for future improvements such as large-scale datasets, attention mechanisms and their rail-time implementation in manufacturing.
