An inspection network with dynamic feature extractor and task alignment head for steel surface defect

High-precision identification and real-time localization for irregular-shaped steel surface defects are crucial for shipbuilding quality control. Although traditional lightweight networks enable real-time defect inspection, the incurred model cannot achieve precise inspection for the defects with large variations in aspect ratios of ship plates. This paper proposes a lightweight inspection network with a dynamic feature extractor and task alignment detection head (INDT) for multi-class steel plate surface defects to address this obstacle. A dynamic structure expansion training strategy based on a re-parameterization multi-branch block is constructed to achieve real-time inspection containing multi-scale information. Furthermore, fed with multi-scale information, the task alignment head with a preprocess for multi-task to concentrate task-oriented features into specific channels. Besides, a soft-weighted sample assignment algorithm with dynamic priors to irregular defects is developed to supervise high-precision model training. The experiments show that the INDT achieves higher precision among all the benchmark methods with lossless accelerating inference.