Cad-Transformer: A CNN–Transformer Hybrid Framework for Automatic Appearance Defect Classification of Shipping Containers

In traditional port operations, manual visual inspection by tally personnel is standard for detecting defects in shipping containers during unloading. However, the fast pace of operations, distractions, poor visibility, and adverse weather often result in errors and delays. To address these issues, we propose the Cad-transformer, a convolutional neural network (CNN)–transformer hybrid framework for automatic container defect classification. The framework includes three components: encoding, upsampling, and decoding. The encoding phase simulates missing parts by randomly deleting 60% of the original image. Upsampling enhances mask resolution, and the decoding phase predicts and fills in missing areas using a transformer-based self-supervised approach. Additionally, the CNN feature extraction method focused on visible patches (CFE-VP) module focuses on key local mask features to accelerate learning. Experimental results show that Cad-transformer outperforms existing methods in container damage detection for smart ports, marking the first use of a CNN-transformer hybrid in this domain.