Research On Aphid Detection Algorithm Based On YOLO

Pests are one of the main threats to agricultural production.Major economic crops such as wheat,corn,rape and cotton are often affected by pests,which causes huge economic losses every year.Therefore,the detection and control of pests is essential.The traditional detection methods are usually identified by experts or farmers based on the external shape characteristics of the pest and the existing records,which is low in detection accuracy,high in false detection rate,and labor-intensive.With the development of deep learning technology,many studies on pest detection based on computer vision technology have been gradually developed.In this paper,we take aphids as the research object and deep learning techniques as the theoretical basis,focusing on the use of convolutional neural networks,attention mechanisms and ASPP feature fusion techniques to detect aphids.The main research work of this thesis is as follows:(1)Introducing sc SE Attention and Spatial Attention to emphasize feature information in images in spatial and channel dimensions and suppress background information interference.The sc SE Attention module is embedded into the back segment in the backbone network,in addition,the Spatial Attention module is deeply embedded inside the Res Unit of the CSP module to balance the efficiency increase and computational redundancy,so as to obtain the best model performance.In addition,the size of anchors is reduced to fit the size of the aphid targets.A set of small anchors is used to replace a set of original anchors to achieve more accurate detection.(2)To address the problem of large relative size differences among aphid individuals in the deeper layers of the network and the problem of large aspect ratio differences,the ASPPFusion module is proposed and the loss function is replaced.The dilation rate of the 4-way dilated convolution inside ASPP is redesigned,and the corresponding feature map size is adapted according to the size of the combination of dilation rate to locate the best embedding position in the backbone network.In addition,the four-way multi-scale context information inside ASPP is fused to balance coarse and fine-grained features to achieve the best detection efficiency.The CIo U loss function is replaced with the EIo U loss function.Compared with CIo U,the penalty term of EIo U loss function splits the height and width of the target and calculates them separately instead of calculating only its aspect ratio,which enables the anchor to fit the aphid target more quickly.Experiments show that the method proposed in this paper can effectively improve the detection accuracy of the object detection model on the aphid dataset.