| 摘 要: 针对计算资源不足的复杂环境下苹果识别效果不佳的问题,提出了一种基于改进 YOLOv11的 CFF-YOLO 轻量化苹果检测算法。首先,将主干网络 C3k2的Bottleneck替换为FasterBlock模块,该模块通过部分卷积(PConv)压缩输入通道至1/4,使计算量(GFLOPs)和内存访问量(MAC)分别降低至标准卷积的1/16和1/4,显著减少了冗余计算;其次,采用FocalModulation模块替换空间池化金字塔(SPPF)结构,增强模型对复杂场景中小目标的感知与定位能力,提升模型检测的精确率;最后,在颈部网络结构中融入轻量化跨尺度特征融合模块 CCFM,通过特征融合技术将多尺度特征进行整合,提升模型对尺度波动的鲁棒性,并强化小目标的检测性能。实验结果表明,CFF-YOLO 模型与基准 YOLOv11模型相比,精确率上涨了2.6%,召回率上涨了8.3%,参数量和计算量分别下降了32.1%和29.2%,模型的大小减小了27.6%。 |
| 关键词: 轻量化 FasterBlock FocalModulation CCFM |
|
中图分类号: 126;TP391.41
文献标识码: A
|
|
| CFF-YOLO: A Lightweight Apple Detection Algorithm Based on Improved YOLOv11 |
|
HAN Yi, WEI Linjing
|
(College of Information Science and Technology, Gansu Agricultural University, Lanzhou 730070, China)
1771605064@qq.com; 916277964@qq.com
|
| Abstract: To address the challenge of effectively recognizing apples in complex environments with limited computational resources, a lightweight apple detection algorithm called CFF-YOLO based on an improved YOLOv11 is proposed. First, the Bottleneck in the backbone network C3k2 is replaced with the FasterBlock module, which compresses the input channels to 1/4 through partial convolution ( PConv). This reduces the computational cost(GFLOPs) and Memory Access Cost ( MAC) to 1/16 and 1/4 of standard convolution, respectively, significantly decreasing redundant computations. Second, the Spatial Pyramid Pooling Fast (SPPF) structure is replaced with the Focal Modulation module to enhance the model’s perception and localization capabilities for small targets in complex scenes, thereby improving detection accuracy. Finally, a lightweight Cross-Scale Convolutional Feature Fusion Module(CCFM) is integrated into the neck network, which integrates mult-i scale features through feature fusion techniques to
enhance the model’s robustness to scale variations and strengthen small target detection performance. Experimental results show that compared to the baseline YOLOv11 model, the CFF-YOLO model achieves a 2.6% increase in precision, a 8.3% increase in recall, and reductions of 32.1% in parameters, 29.2% in computational cost, and 27.6% in model size. |
| Keywords: lightweight FasterBlock FocalModulation CCFM |
