Authors

Abstract

To accurately detect minor surface defects on steel components of agricultural machinery in complex environments, this study proposes an improved detection model, CGC-YOLOv11n, based on the YOLOv11n architecture. First, Converse2D reverse convolution is integrated into the C3k2 module to enhance fine-grained feature representation for subtle and blurred defects. Second, the GESA module replaces C2PSA, leveraging dynamic sparse attention to strengthen multi-scale aggregation and focus on key target cues. Third, a Coordinated Detail-Preserving Contextual Fusion (CDPCF) module—an innovative extension of DPCF—is embedded in the neck, employing adaptive content-aware gating to synergistically balance high- and low-resolution features. Experimental results demonstrate that the CGC-YOLOv11n achieves a precision of 74.85%, a recall of 74.16%, and a mean average precision (mAP@0.5) of 79.81%, representing an improvement of 1.83 percentage points over the baseline YOLOv11n. Across another dataset, CGC-YOLOv11n delivers superior detection performance. The improved model exhibits robust capabilities in agricultural machinery maintenance, providing technical support for reliable agricultural operations.

Abstract in Chinese

为在复杂环境中精确检测农机钢制部件的微小表面缺陷，本研究基于YOLOv11n架构提出改进检测模型CGC-YOLOv11n。首先, Converse2D反卷积集成到C3k2模块中，以增强对细微模糊缺陷的精细特征表征能力。其次GESA模块替代C2PSA，通过动态稀疏注意力机制强化多尺度聚合和对关键目标线索的关注能力，精准聚焦关键目标特征。第三，在颈部嵌入创新性扩展模块CDPCF（协调细节保留上下文融合），该模块基于DPCF模块，采用自适应内容感知门控机制协同平衡高/低分辨率特征。实验结果表明，CGC-YOLOv11n模型实现74.85%的精确率、74.16%的召回率及79.81%的平均精确率（mAP@0.5），较基线YOLOv11n提升1.83个百分点。在另一数据集测试中，CGC-YOLOv11n展现出更精准的检测效果。该优化模型在农机维护领域表现出强健性能，为可靠农业作业提供了技术保障。