Lightweight Deep Learning Model for TFT-LCD Pixel Defect Detection Using MobileNetV2

Abstract

The proposed methodology aims to achieve high classification accuracy while maintaining computational efficiency and minimizing inference latency, making it highly suitable for real-time industrial inspection scenarios. This study evaluated five traditional transfer learning models (VGG16, ResNet50, MobileNetV2, Xception, and InceptionV3) for comparison in pixel defect detection. MobileNetV2-based implementation demonstrated a competitive average test accuracy of 0.9916, outperforming several baselines while maintaining significantly lower computational complexity. Based on our investigation, MobileNetV2 offers an optimal balance between accuracy and efficiency, making it a preferred choice over heavier architectures such as VGG16 and ResNet50, which, despite achieving comparable accuracy of 1.0, incur substantially greater computational overhead and longer inference times. These results underscore the suitability of MobileNetV2 as a lightweight and reliable alternative for automated visual quality inspection in display manufacturing, particularly where real-time performance and operational efficiency are critical.