基于改进YOLOv5s的轻量化储粮害虫检测方法

Lightweight Grain Storage Pest Detection Method Based on Improved YOLOv5s

粮食的安全储藏关系着国计民生,在粮食储藏过程中快速、精准检测储粮害虫具有重要意义。目前传统储粮害虫检测手段多数是通过人工检测,耗时、误判而且成本高。实验通过深度学习技术,提出一种轻量化YOLOv5的害虫检测方法。首先,通过优化锚框参数;其次,将轻量化的MobileNetv3中的注意力模块替换成更高效的ECA,然后把改进后的MobileNetv3取代YOLOv5s的主干网络;最后,用加权双向特征金字塔Bi-FPN结构替换YOLOv5s中的特征金字塔结构。在自制储粮害虫数据集上训练结果显示,改进后的算法模型平均准确率达到97.1%,FPS提高至91,模型计算量仅有原来的11.6%。实验表明,与原始YOLOv5s相比,改进后的算法在储粮害虫小目标检测上精度更高、速度更快、模型更轻量化。

The safe storage of grain is related to the national economy and people′s livelihood,and it is of great significance to quickly and accurately detect stored grain pests during grain storage.At present,most of the means of detecting pests in food storage are manual detection,time-consuming,misjudged and costly.In this research,a pest detection method for lightweight YOLOv5 was proposed through deep learning technology.Firstly,the anchor box parameters were optimized.Secondly,the attention module in the lightweight MobileNetv3 was replaced with a more efficient ECA,and then,the YOLOv5s backbone network was replaced with the improved MobileNetv3.Finally,the feature pyramid structure in YOLOv5s was replaced with a weighted bidirectional feature pyramid Bi-FPN structure.The training results on the self-made grain storage pest dataset indicated that the average accuracy of the improved algorithm model reached 97.1%,the FPS was improved to 91,and the model calculation amount was only 11.6%.Experiments indicated that,compared with the original YOLOv5s,the improved algorithm had a higher accuracy,faster speed and lighter model in the detection of small targets of stored grain pests.