Defect transformer: An efficient hybrid transformer architecture for surface defect detection

Surface defect detection is an extremely crucial step to ensure the quality of industrial products. Nowadays, convolutional neural networks (CNNs) based on encoder–decoder architecture have achieved tremendous success in various tasks. However, intrinsic locality convolution prevents them from modeling long-range interactions explicitly, making it difficult distinguish pseudo-defects cluttered backgrounds. Recent transformers are especially skilled at learning global image dependencies, but with limited local structural information for refined location. To overcome above limitations, we incorporate CNN and transformer into efficient hybrid detection, termed Defect Transformer (DefT), capture non-local relationships collaboratively. Specifically, encoder module, a stem block firstly adopted retain more spatial details. Then, patch aggregation blocks used generate multi-scale representation four hierarchies, each followed by series DefT blocks, which respectively include locally position-aware position encoding, lightweight multi-pooling self-attention model contextual good computational efficiency, feed-forward network feature transformation further learning. Finally, simple effective decoder module constructed gradually recover details skip connections encoder. Extensive experiments three datasets demonstrate superiority efficiency our method compared other deeper complex CNN- transformer-based networks.