基于MS-YOLO v7的多尺度稻飞虱识别分类方法

Multi-scale Rice Planthopper Image Recognition and Classification Based on Lightweight MS-YOLO v7

智能虫情测报灯下害虫的精准识别和分类是实现稻田虫情预警的前提,为解决水稻害虫图像识别过程中存在分布密集、体态微小、易受背景干扰等造成识别精度不高的问题,提出了一种基于MS-YOLO v7(Multi-Scale-YOLO v7)轻量化稻飞虱识别分类方法。首先,采用稻飞虱害虫诱捕装置搭建稻飞虱害虫采集平台,获取的稻飞虱图像构成ImageNet数据集。然后,MS-YOLO v7目标检测算法采用GhostConv轻量卷积作为主干网络,减小模型运行的参数量;在Neck部分加入CBAM注意力机制模块,有效强调稻飞虱区别度较高的特征通道,抑制沉冗无用特征,准确提取稻飞虱图像中的关键特征,动态调整特征图中不同通道的权重;将SPPCSPS空间金字塔池化模块替换SPPFS金字塔池化模块,提高网络模型对各分类样本的特征提取能力;同时将YOLO v7模型中的SiLU激活函数替换为Mish激活函数,增强网络的非线性表达能力。试验结果表明,改进后的MS-YOLO v7在测试集上的模型平均精度均值(Mean average precision,mAP)为95.7%,精确率(Precision)为96.4%,召回率(Recall)为94.2%,与Faster R-CNN、SSD、YOLO v5、YOLO v7网络模型相比mAP分别提高2.1、3.4、2.3、1.6个百分点,F1值分别提高2.7、4.1、2.5、1.4个百分点。改进后的模型内存占用量、参数量、浮点运算数分别为63.7 MB、2.85×107、7.84×1010,相比YOLO v7模型分别缩减12.5%、21.7%、25.4%,MS-YOLO v7网络模型对稻飞虱种间害虫均能实现高精度的识别与分类,具有较好的鲁棒性,可为稻田早期稻飞虱虫情预警提供技术支持。

Accurate identification and classification of pests under intelligent insect monitoring and reporting lights are the prerequisite for realizing early warning of rice insect situation.In order to solve the problems in image recognition of rice pests,such as dense distribution,small body size and susceptibility to background interference,the recognition accuracy is not high.A lightweight MS-YOLO v7(Multi-Scale-YOLO v7)based classification method for rice fly identification was proposed.Firstly,a rice planthopper pest collection platform was built with a migratory pest trapping device,and the images of rice planthopper were obtained to form the ImageNet dataset.Then the MS-YOLO v7 object detection algorithm used GhostConv lightweight convolution as the backbone network to reduce the number of parameters for model operation.CBAM attention mechanism module was added to Neck to effectively emphasize the highly differentiated feature channels of rice planthopper,suppress redundant and useless features,accurately extract key features of rice planthopper images,and dynamically adjust the weights of different channels in the feature map.SPPCSPS spatial pyramid modulewas replaced by SPPFS pyramid module to improve the feature extraction ability of the network model.At the same time,SiLU activation function was replaced by Mish activation function in YOLO v7 model to enhance the nonlinear ability of the network.The test results showed that the mean average precision(mAP),precision(96.4%)and recall(94.2%)of the improved MS-YOLO v7 on the test set were 95.7%,96.4%and 94.2%,respectively.Compared with that of Faster R-CNN,SSD,YOLO v5 and YOLO v7 network models,mAP was improved by 2.1 percentage points,3.4 percentage points,2.3 percentage points and 1.6 percentage points,respectively,and the balance score F1 was improved by 2.7 percentage points,4.1 percentage points,2.5 percentage points and 1.4 percentage points,respectively.The memory occupation,number of parameters,and number of floating-point operations of the improved model were 63.7 MB,2.85×107,and 7.84×1010,respectively,which were scaled down by 12.5%,21.7%,and 25.4%compared with that of the YOLO v7 model.The MS-YOLO v7 network model can realize highprecision identification and classification of interspecific pests of rice fly,with good robustness,and it can be used to realize the technical support for early warning of rice fly pest in paddy fields.