Traditional defect detection methods for nuclear power plant concrete structures rely on inspectors, which are subjective and labor-intensive. Enhancing visual inspection with computers is crucial since concrete defects pose safety risks and can lead to severe structural failures. This paper investigates a lightweight Convolutional Neural Network (CNN) approach to improve the detection and classification of these defects. By leveraging a multi-head MobileNet architecture and incorporating data augmentation through generative modeling, we aim to improve prediction accuracy and efficiency. Our proposed method offers a more reliable and safer alternative to conventional inspection techniques, thus guaranteeing better maintenance and safety of nuclear power plant infrastructure.