An efficient industrial defect detection based on hybrid residual attention with modified generative adversarial network and convolutional neural network model

Abstract

Detecting and classifying industrial defects continues to pose significant issues in Industry 4.0 era, principally due to constraints in managing data scarcity, variability in fault characteristics, and the inadequacy of traditional models to adapt to different and dynamic situations. Contemporary approaches frequently encounter difficulties in producing dependable synthetic data, effectively extracting essential features, and attaining robust performance in industrial applications. This paper presents a hybrid residual attention generative adversarial network with convolutional neural networks (RAtGAN-CNN) model to address these constraints. The RAtGAN-CNN framework combines residual blocks and attention mechanisms with a generative adversarial network to produce high-quality synthetic samples that replicate the complex distributions of real defects. This approach effectively addresses data scarcity and is trained concurrently with a discriminator through adversarial learning, thereby enhancing data diversity and reducing overfitting in situations with limited labeled data. The lightweight design guarantees appropriateness for real-time industrial applications, fulfilling the demands of computationally limited situations. The model employs a lightweight convolutional neural network (CNN) that utilizes a modified residual block to boost feature extraction, while its attention mechanism concentrates on critical defect areas to improve detection accuracy. These methodologies empower the RAtGAN-CNN to operate effectively across many datasets and settings, particularly excelling in situations with sparse or highly variable input data. The framework is evaluated on a binary image classification dataset of industrial casting defects, attaining a competitive accuracy above 99% on the validation set, with a lightweight model size of 12.5 MB and an average inference time of 18.5 ms per image on a single GPU. Metrics including precision, recall, and F1-score illustrate the approach’s robustness, underpinned by thorough evaluation via confusion matrices and loss-accuracy curves.