ALADA : A lite automatic data augmentation framework for industrial defect detection

Industrial defect detection is a critical and challenging task in the quality control of manufacturing production. Competent in feature extraction and pattern recognition, deep learning shows great power for classifying and locating defects in industrial products. However, insufficient data records and diverse categories of defects restrict the detection accuracy of data-driven neural networks. One direction to solve such a problem is data augmentation, which aims at generating synthetic copies from existing data and improving the generalizability of detection models. However, confirming a suitable augmentation policy involves either human experience or substantial experiments for augmentation parameters. To address these challenges, in this work, a lite automatic data augmentation (ALADA) framework is proposed to jointly optimize the data augmentation policies and the neural network for industrial defect detection. First, a lite search space is formulated to efficiently sample augmentation policies and generate augmented images for joint optimization. To reduce the hyperparameter tuning efforts for retraining with searched policies, a three-step bi-level optimization scheme is proposed to replace the retraining strategy and update the model and augmentation parameter alternately. To solve the non-differentiable problem in the joint optimization scheme, policy gradient sampling is implemented to estimate the gradient flow efficiently. Experimental results on three industrial defect detection datasets, namely, Tianchi-TILE, GC10-DET, and NEU-DET, reveal that our proposed automatic augmentation framework outperforms the state-of-the-art augmentation methods and effectively improves the accuracy of the baseline defect detection model. The proposed ALADA scheme also alleviates the missed detection of defects in four practical industrial circumstances: textured background, uneven brightness, low contrast, and intraclass difference. © 2023 Elsevier Ltd. All rights reserved