Abstract: In order to ensure the accuracy and rapidity of the embedded weeding robot in the corn field, a real-time target detection algorithm based on GBC-Yolov5s is proposed. First, the combination of the 1×1 convolution and depth-separable convolution is used to replace the traditional convolution, which reduces the redundant features generated by the backbone network without changing the size of the output feature map. Secondly, a bidirectional feature fusion network (S-BiFPN) network is designed to enhance the ability of feature extraction, which can make full use of different scale features to improve the speed of weed detection and combine the multi-channel structure with the self-attention mechanism to enhance the attention of small targets by compressing and reweighting the input features. Finally, MWeed data sets are built for different environments to test the proposed algorithm. The results show that compared with the Yolov5s and Faster RCNN model algorithms, the size of the GBC-Yolov5s model after lightweight is only 3.3 MB, the detection time of the input image (GPU) reaches 15.6 ms, and the average accuracy (mAP) reaches 96.3%, which can effectively improve the target detection speed and recognition accuracy, and provide a theoretical basis for the field of intelligent agricultural weeding.

Key words: YOLOv5s, target identification, model compression, feature fusion

摘要： 为了实现嵌入式除草机器人在玉米田间准确、快速的进行除草工作，提出了一种实时目标检测算法GBC-YOLOv5s。使用1×1卷积和深度可分离卷积的组合替代普通卷积，在不改变输出特征图大小的情况下减少主干网络产生的杂草冗余特征。设计了一种双向特征融合网络（S-BiFPN）增强特征提取能力，充分利用不同尺度的特征提高杂草检测速度，并将多通道结构与自注意力机制结合，通过对输入特征进行压缩与再加权，以加强对小目标的关注度。针对不同的环境构建MWeed数据集进行测试，结果表明，与现有Yolov5s、Faster RCNN等模型方法相比，GBC-YOLOv5s模型轻量化后的大小仅为3.3 MB，输入图像的检测耗时（GPU）达到15.6 ms，平均精度（mAP）达到96.3%，能够有效地提升目标检测速度和识别精度，为农业智能除草领域提供理论依据。

关键词: YOLOv5s, 目标识别, 模型压缩, 特征融合