Abstract: High‑quality goose down is widely used in thermal insulation products due to its excellent loft, thermal insulation performance and softness, and its market value is significantly higher than that of duck down. However, in actual production, different types and qualities of down materials are often mixed, which seriously affects product quality and consumer trust. At present, the identification of goose down and duck down mainly relies on manual visual inspection, which is highly subjective and inefficient, and is prone to misjudgment and missed detection when dealing with down barbule clusters with small scales and complex structures. Although previous studies have attempted to apply traditional machine learning and deep learning methods to down classification, achieving high‑precision and real‑time automatic recognition remains challenging due to the weak visual features of barbules and the high proportion of small targets. To address these issues, this paper proposed a lightweight detection model named FCA‑YOLO based on feature fusion and attention mechanisms for automatic down type identification. First, a dedicated down image acquisition system was designed and constructed to collect high‑resolution grayscale images at a magnification of 914×, and a specialized dataset containing goose and duck down barbule clusters was established with manual annotation. Second, based on the YOLOv4‑tiny architecture, a large‑scale shallow feature layer Feat3 was introduced to preserve more spatial detail of small targets at the early stages of the network, thereby enhancing the detection capability for fine structures such as down barbules. Meanwhile, a convolutional block attention module (CBAM) was embedded to adaptively optimize feature representations from both channel and spatial dimensions, enabling the network to focus more effectively on discriminative regions. Furthermore, an attentional feature pyramid network (AFPN) was incorporated to fuse multi‑scale features, effectively alleviating information loss and semantic inconsistency commonly encountered in traditional feature pyramids during non‑adjacent layer fusion. Extensive experiments were conducted to systematically evaluate the proposed model in terms of detection accuracy, robustness and computational efficiency. Ablation studies demonstrate that each introduced module contributes positively to performance improvement. The complete FCA‑YOLO model achieves a mean average precision (mAP) of 76.14% on the test set, representing an improvement of 8.24% over YOLOv4‑tiny. Comparative experiments with YOLOv4, YOLOv4‑tiny, Faster R‑CNN, YOLOv4‑MobileNetV3, RTDETR‑ResNet50, YOLOv8n and YOLO12n further verify that FCA‑YOLO achieves a favorable balance between detection accuracy and computational complexity, particularly excelling in small‑scale goose down barbule cluster detection tasks. In single‑image recognition, the model attains an overall classification accuracy of 98.4% for goose down, duck down and unknown down, with an average inference time of only 0.01 s per image. The experimental results indicate that FCA‑YOLO significantly enhances fine‑grained down recognition performance while maintaining a compact architecture and low computational cost, demonstrating strong potential for real‑time detection and embedded deployment applications.

Key words: down type recognition, YOLOv4?tiny, large?scale shallow feature layer, CBAM, AFPN

摘要： 针对现有鹅绒、鸭绒人工识别分类效率低、易漏检误判以及传统识别算法鲁棒性差、识别准确低的问题,提出了一种轻量化深度网络模型 FCA-YOLO。 首先,在 YOLOv4-tiny 网络的颈部结构中引入大尺度浅层特征层 Feat3,以提升对羽绒结点群等小目标的识别精度;同时引入卷积注意力模块,从通道和空间两个维度对特征信息进行优化;最后,采用渐进式特征金字塔网络融合不同尺度的特征。 结果表明:所提出的 FCAYOLO 模型在测试集上的 mAP 为 76. 14%,相比原模型提升了 8. 24%;模型体积仅增加 4. 9 MB,同时 GFLOPs仅增加 0. 3,参数量降低 0. 37 M。 在单幅图像识别任务中,FCA-YOLO 对鹅绒、鸭绒和未知绒的总体识别正确率 98. 4%,识别时间仅为 0. 01 s。 研究模型在保证识别精度的同时兼顾了模型轻量化特性,具有一定的工程应用潜力。

关键词: 羽绒种类识别, YOLOv4-tiny, 大尺度浅层特征层, CBAM, AFPN