A Comprehensive Investigation on Convolutional Neural Network-Based Product Defect Detection in Industrial Applications

Abstract

Ensuring high product quality has become increasingly important in modern manufacturing. Traditional manual inspection and basic machine vision systems often perform not well in accuracy, efficiency, and robustness, especially in complex industrial environments. Convolutional Neural Networks (CNNs), which is a technology in deep learning specialized for visual data, have shown great potential in product defect detection. A broad range of recent CNN-based methods for product defect detection is systematically reviewed in this article by classifying existing methods into three major types: baseline CNN models, attention-based models, and hybrid CNN frameworks. Baseline models such as MobileNet, ResNet, and EfficientNet focus on improving accuracy and computational efficiency. Attention-based models incorporate spatial and channel attention mechanisms to localize subtle defects better. Hybrid models combine CNNs with techniques like Random Forests or adaptive fusion strategies to improve robustness across varied defect types and product conditions. In addition to summarizing technical advancements, this paper highlights three critical challenges that hinder real-world deployment: (1) lack of model interpretability, (2) limited applicability, and (3) difficulty in deploying models on resource-constrained edge devices. To address these, this paper further discusses promising solving directions, including the integration of domain knowledge and expert systems, domain adaptation and generalization strategies, and model optimization methods like model pruning, reduced-precision quantization, and student–teacher distillation frameworks. Overall, this review provides a systematic overview, offering valuable insights for researchers and engineers, and explores future research paths for enhancing CNN-based defect detection in manufacturing.