LFA-DEIM: A lightweight feature aggregation framework with improved DETR matching for real-time steel plate defect detection

Abstract Steel plate surface defect detection remains challenging due to the trade-off between small target detection accuracy and computational efficiency, background texture interference, and multi-scale defect variations. We present LFA-DEIM, a lightweight feature aggregation framework with improved DETR with improved matching (DEIM) for real-time defect detection. The method employs a biologically-inspired hierarchical kernel convolution strategy combining large-field perception with small-region focusing, achieving efficient multi-scale feature extraction. A spatial-frequency dual-domain mechanism integrates edge gradient enhancement with frequency-selective filtering to discriminate periodic textures from aperiodic defects. An adaptive deformable sampling strategy with learnable offsets enables dynamic receptive field adjustment across defect scales. Experiments on NEU-DET and GC10-DET datasets demonstrate mean average precision@0.5 (mAP@0.5) of 83.5% and 81.9% respectively, improving 4.7 and 4.3 percentage points over baselines. Model size reduced to 15.1 MB with 24.8 GFLOPs computational cost, achieving 104.2 frames per second (FPS) on personal computer (PC) and 56.5 FPS on NVIDIA Jetson Nano platforms. The approach successfully balances detection accuracy, inference speed, and deployment feasibility for industrial real-time inspection.