| Navel oranges,as a characteristic agricultural product of the Ganan region,play a vital role in the surrounding social and economic development,and navel orange pests are a key factor in determining the yield and quality of navel oranges.Therefore,rapid and accurate detection of navel orange insect pests is of great research importance.The task of detecting pests in navel oranges is influenced by factors such as the complex natural environment,which leads to increased difficulty in detection,high rates of false and missed detections and low detection efficiency.Thus,efficient detection of navel orange pests has turned out to be a hot topic of concern for the computer vision industry.Compared to traditional target detection methods,deep learning-based target detection methods have higher detection accuracy and stronger adaptability,and are also widely used in the agricultural field.The classical target detection algorithm YOLOv5 s model is gradually emerging in the field of target detection due to its small size and low deployment cost.While navel orange pest images have small targets and complex backgrounds,YOLOv5 s processing will result in false detection,leading to a decrease in accuracy.In this study,we propose a target detection algorithm for navel orange pests based on an improved YOLOv5 s model,using nine common navel orange pests as the research object.The main work is as follows.(1)The YOLOv5 s algorithm was applied to the navel orange pest dataset for initial training and validation.First,the navel orange pest dataset was constructed,and nine types of navel orange pest images were collected and then annotated.For the problem of small sample size and poor training effect of the navel orange pest dataset,several data enhancement methods were combined to enrich the dataset,prevent overfitting of the training and improve the robustness of the network model.A complex background dataset of 18,000 navel orange pest images was finally obtained.The experimental environment for YOLOv5 s navel orange pest detection was then constructed,and preliminary judgments were made based on the experiments,and the accuracy of YOLOv5 s for navel orange pest detection needs to be improved.(2)A stage-by-stage improvement of the YOLOv5s-based pest detection algorithm for navel oranges is proposed.To solve the problems of inadequate extraction of complex target features and large pest localization bias in navel orange pest detection,YOLOv5 s was improved in three aspects by combining different parts of the YOLOv5 s model: the CBAM attention mechanism was added to the backbone network to improve the linkage of target features in channels and space and enhance the network’s attention to key features,thereby improving the feature extraction capability for navel orange The Si LU activation function in the YOLOv5 s model was replaced with the FRe LU activation function,which is more suitable for image data,in order to improve the feature expression capability of the activation function,so that it has the capability of adaptively acquiring local contextual information of the image and expanding the reception domain of the model.The experimental results show that the average mean accuracy m AP values of all three improvements have achieved a certain degree of improvement.Finally,a test set of 450 images was built to evaluate the model performance.Based on the pest prediction images,it was shown that the multi-stage improved YOLOv5s-CFE algorithm has greatly improved the problems of false detection and missed detection of pest targets and inaccurate positioning of prediction frames.(3)A lightweight navel orange pest detection algorithm based on YOLOv5s-CFE is proposed.A lightweight network model YOLOv5 sCFEG is proposed to meet the requirement of lightweight network model for later network application deployment.detection algorithm,while reducing the computational complexity and compressing the model parameters to facilitate model deployment at a later stage. |
PDF Full Text Request