An Integrated AI Automating Inspection Model for Surface Defect Segmentation and Classification in Construction Domain
TL;DR
This work gives unified Deep Learning model which evaluates on one of the public benchmark datasets from Kaggle which consists of sub folders of Magnetic-Tile Defect, DeepPCB, DBCC bridge cracks and CrackForest and is trained from scratch without pre-training and employs objective which is imbalanced.
Abstract
Unlike stone age, humans no longer live in caves or mud houses but rather shifted to modern civilization setups which include - buildings, bridges and industrial facilities as their environment. And with time these structures and components develop surface defects like cracks, peeling, spalling, algae, staining and manufacturing flaws. If frequently, these structures are left untreated, then it may affect human safety and threaten serviceability. For inspection purposes, only relying on manual visual examination will lead to a slow process and some parts may be left unchecked. This will add pressure on pockets. Moreover, this method is slow, subjective, costly, error-prone and unsafe for surfaces which are inaccessible. Overall, these factors motivate automated computer vision inspection. However, there exists some constraints like limited, imbalanced datasets and inspection spans of various distinct types of tasks. This work gives unified Deep Learning model which evaluates on one of the public benchmark datasets from Kaggle which consists of sub folders of Magnetic-Tile Defect, DeepPCB, DBCC bridge cracks and CrackForest. In this paper multi-class classification, binary classification and pixel-level segmentation are considered. The model couple’s residual backbone with classification, which is regularized head and symmetric U-Net. Overall is trained from scratch without pre-training and employs objective which is imbalanced (class-weighted, positive weighted BCE with Dice, label-smoothed cross entropy). This framework attains F1 score of 0.9945 for DeepPCB dataset. The novelty lies in its approach, as its single, pre-training free pipeline which spans three task paradigms with the issue of handling imbalance and format agnostic ingestion and in real honest finding that scale of data, more than architecture and govern effectiveness. Future work could focus on adding pre- trained backbones, multi-seed evaluation and board- disjoint and then end to end detection.