Co-training encoder-based transformer method for steel surface defect detection

As a critical component of industrial production, quality inspection of steel plays an essential role. However, the prevalent convolutional neural network–based detection methods are constrained by their reliance on manual post-processing operations, such as non-maximal suppression, and their limited capacity to model long-range dependencies within images. On the other hand, transformer-based detection algorithms, although promising, have shown relatively weaker performance in defect detection. This is primarily due to challenges such as difficulty in loss convergence and low detection accuracy for small objects. To address these issues, we present a detection transformer architecture called co-training encoder-based detection transformer (CE-DETR). Our approach features a co-training encoder structure, which enhances both convergence speed and detection accuracy by leveraging comparative learning with pretrained encoder features. Furthermore, a weighted bidirectional feature pyramid network is integrated post-backbone, providing the encoder with input features that possess both rich semantic information and precise spatial localization. Furthermore, we introduce an improved hybrid dynamic query selection mechanism, based on IoU-aware, to supply the decoder with higher-quality initial object queries, thereby further optimizing detection performance. Experimental results demonstrate that CE-DETR achieves a detection accuracy of 78.6 [email protected] on NEU-DET, surpassing other compared detection algorithms. Moreover, the proposed method exhibits strong generalization capabilities in real-world industrial scenarios, as evidenced by its detection accuracy of 71.3% [email protected] on the GC 10-DET dataset. The codes are available at https://github.com/903770351/CE-DETR.git