基于改进轻量化YOLOv5s的万寿菊识别方法

Marigold recognition method based on improved lightweight YOLOv5s

摘要:

针对传统识别网络结构冗余且普适性低，导致推理过程复杂、效能低下，难以满足密集生长的万寿菊检测需求的问题，提出了一种改进的轻量化YOLOv5s万寿菊识别方法，并通过适宜的数据集增强技术为识别方法提供可靠的数据支撑；该方法采用ShuffleNet V2替代CSPDarknet53作为主干网络，并引入SimAM注意力机制，以减小网络规模并提升密集目标检测效能；颈部网络采用Slim-neck结构，结合GSConv和VoV-GSCSP模块提升密集目标的特征提取效率；训练过程中，使用WIoU误差函数并通过Soft-NMS动态调整边界框，以增强网络的泛化能力；算例分析及产地实测结果表明，所改进的轻量化YOLOv5s网络的平均精度均值较现有常用模型提高了3.00 %，参数量减少了6.44 MB，每秒浮点运算次数减少了14.70 G，模型体积减少了12.24 MB，每秒帧数增加了47.19 帧，且网络鲁棒性强，极大地降低了其应用与推广成本。

Abstract:

Aiming at the problem that the traditional recognition network structure is redundant and has low universality, which leads to complex reasoning process and low efficiency, and it is difficult to meet the needs of dense growth marigold detection, an improved lightweight YOLOv5s marigold recognition method is proposed, and reliable data support is provided for the recognition method through appropriate data set enhancement technology. This method uses ShuffleNet V2 instead of CSPDarknet53 as the backbone network, and introduces the SimAM attention mechanism to reduce the network size and improve the detection efficiency of dense targets. The neck network adopts the Slim-neck structure, and combines the GSConv and VoV-GSCSP modules to improve the feature extraction efficiency of dense targets. During the training process, the WIoU error function is used and the bounding box is dynamically adjusted through Soft-NMS to enhance the generalization ability of the network. The results of example analysis and field measurement show that the average accuracy of the improved lightweight YOLOv5s network is 3.00 % higher than that of the existing commonly used model, the parameter amount is reduced by 6.44 MB, the number of floating-point operations per second is reduced by 14.70 G, the model volume is reduced by 12.24 MB, and the number of frames per second is increased by 47.19 frames. The network has strong robustness, which greatly reduces its application and promotion costs.