Research on Intelligent Detection of Container Surface Damage Based on Deep Residual Networks

Abstract

Addressing the issues of low efficiency and high subjectivity in container damage detection during automated port operations, this study employs deep residual networks to construct a high-precision system for determining the presence or absence of container damage. Based on a dataset of 3,713 images, the approach achieves the goal of extracting key damage features from complex port backgrounds by converting original YOLO annotations into binary classification labels. Targeted data augmentation strategies, including geometric and illumination transformations, were employed to address challenges such as reflections, shadows, and multi-scale damage in real-world scenarios. The model architecture incorporates three residual blocks, global average pooling, and a Sigmoid output layer, with skip connections effectively mitigating gradient vanishing during deep training. Experiments employed 3-fold hierarchical cross-validation and ensemble learning strategies, incorporating label smoothing and early stopping mechanisms to optimize training. Results demonstrate the model achieves 99.39% accuracy, 100% precision, and 99.39% recall on the validation set, while maintaining only 2.81 million parameters. This significantly enhances detection robustness and stability, providing reliable technical support for smart port development.