GMFC-YOLO11n: a lightweight object detection network for microscopic goose and duck down identification

Abstract

Down materials are widely used in high-end apparel fields, and their quality grading is crucial to industrial development. However, goose down and duck down exhibit highly similar microstructures with inherent weak textures. Traditional manual detection suffers from low efficiency and strong subjectivity, hindering automated industrial inspection. This paper proposes an improved lightweight object detection model named GMFC-YOLO11n. Based on YOLO11n, systematic structural optimizations are implemented: the Spatial and Channel Reconstruction Convolution (SCConv) module is introduced in the backbone network to enhance spatial and channel feature reconstruction capability; the C3k2_FasterBlock is adopted in the feature extraction stage to reduce computational complexity and improve multi-scale feature representation; Gated-Spatial Convolution (GSConv) is embedded in the feature fusion stage to boost cross-channel information interaction. The mixed local channel attention (MLCA) attention mechanism is integrated to strengthen the joint modeling of local textures and global semantics, thereby improving the model’s perception of fine-grained down microstructures. Experimental results on the self-established microscopic down dataset Microscopic Feather Identification Dataset-70 (MFID-70) show that the proposed method achieves Precision of 0.631 and Recall of 0.637, with respective improvements of 3.27% and 4.09% compared with the original YOLO11n. In terms of lightweight indicators, GMFC-YOLO11n has 2.41 M parameters and 6.2 G floating-point operations (FLOPs), with a model size of 9.6 MB. Compared with the baseline YOLO11n, its parameter volume and computational overhead are reduced by approximately 6.95% and 3.13%, with an inference speed of 129.25 FPS. The model maintains a balance between detection accuracy and computational efficiency.