SFM-YOLO: a small target detection and identification method based on spatial frequency characteristics and multi-direction mamba under low contrast condition

Abstract The structural integrity of the aero-engine is crucial for ensuring aviation safety. As one of the core components of the aero-engine, any defects in the blades may lead to engine damage and pose significant risks to flight safety. However, the small scale and low contrast of blade surface defects hinder effective feature extraction, while missed detections and inaccurate localization further reduce identification accuracy. To address these challenges, this paper proposes an improved lightweight small target detection framework, named SFM-YOLO, to tackle the challenge of detecting tiny surface defects on aero-engine blades. First, this paper introduces a Spatial-Frequency Joint Attention Mechanism, which integrates spatial-domain and frequency-domain features through cross-indexing and adaptive learning, thereby enhancing the localization accuracy of aero-engine blades and reducing missed detections. Furthermore, a Spatial-Frequency Dynamic Convolution module is designed, which employs dual-frequency-domain decomposition with multi-branch dynamic convolution enhancement to improve the detection capability of small targets under low-contrast conditions. Finally, this paper proposes a lightweight Multi-Scale Direction Mamba module, which utilizes multi-scale state space modeling and small target pyramid enhancement to improve global feature representation and strengthen edge details, thereby further enhancing the detection accuracy of small targets. Experimental results demonstrate that SFM-YOLO outperforms existing state-of-the-art methods on the AeBST and NEU-DET datasets. Under limited computational cost, it achieves a 5.2% improvement in precision, a 2.3% improvement in recall, and a 3.2% improvement in mAP50. The proposed method enables the precise detection and identification of small target defects on aero-engine blades, even under limited hardware resources, making it more suitable for deployment in engineering applications and practical aero-engine inspection systems.