Research On Lightweight Object Detection Algorithms For Pest And Disease Recognition

With the increasing processing power of Graphics Processing Units(GPUs),deep learningbased object detection techniques have been widely applied in the field of crop pest and disease recognition.Crop pest and disease are one of the most complex,variable,and difficult-toovercome factors that restrict crop growth,causing significant economic losses to agricultural production in our country.Traditional pest and disease identification methods mainly rely on manual identification or machine learning classifiers,which are inefficient and prone to errors.Deep learning technology,on the other hand,provides a more accurate and efficient solution.Although deep learning-based pest and disease detection techniques are becoming increasingly mature,existing deep learning algorithms have large model sizes and high computational requirements,making it costly to deploy them on embedded devices,which is not conducive to practical applications in pest and disease recognition scenarios.Therefore,finding a balance between accuracy and model complexity is particularly important.In order to address the crucial issue of deep learning technology in the industrialization process of agricultural pest and disease recognition,there is an urgent need to design a lightweight object detection model that can be deployed on embedded devices with limited computing resources.This paper focuses on this problem and carries out the following specific work:(1)Building a custom pest and disease dataset.In response to the lack of specific pest and disease datasets in the Guangxi region,this study collected images of cabbage loopers and citrus canker disease in the Wuming District of Nanning City.The dataset was expanded to 1679 images using various data augmentation techniques such as histogram equalization.These images were then labeled using the label Img tool to create a VOC format dataset for all the experiments conducted in this study.(2)Constructing pest and disease recognition models using classical object detection algorithms and designing two lightweight algorithms: Res Net50-Faster R-CNN and Mobile Netv2-SSD.Firstly,classic two-stage object detection algorithms such as Faster R-CNN(VGG16-Faster R-CNN)and one-stage object detection algorithms such as SSD(VGG16-SSD),YOLOX,and YOLOV5 were utilized to build models and conduct experimental testing on the self-built dataset.Secondly,based on the aforementioned research,optimizations were made to the feature extraction networks of the original Faster R-CNN and SSD models,which used VGG16.The optimized Res Net50-Faster R-CNN achieved a 1.6% improvement in m AP and a 70.9% reduction in model size compared to the original algorithm.The Mobile Netv2-SSD achieved a 3% improvement in m AP and an 83.7% reduction in model size compared to the original algorithm.The experimental results demonstrate that designing efficient and concise feature extraction networks can enhance m AP while reducing model size.Finally,a performance comparison analysis was conducted among the six aforementioned models,and the results indicate that the improved Res Net50-Faster R-CNN and Mobile Netv2-SSD showed performance improvements but still exhibited slight differences compared to YOLOV5.(3)Addressing the problem of high computational costs caused by the high utilization of ordinary convolutions in the original YOLOV5 network structure,this study proposed a GCAYOLOV5 algorithm based on Ghost Net and Channel Attention(CA).By incorporating the design principles of Ghost Net and CA attention mechanisms,improvements were made to the backbone and neck networks of YOLOV5.The experimental results showed that the improved GCAYOLOV5 achieved a 0.2% improvement in m AP,a 56.9% reduction in model size,a 59.2%decrease in parameters,and a 51.5% reduction in floating-point operations compared to the original YOLOV5.Additionally,comprehensive performance comparisons were conducted among the six models established in the previous work on the server-side.The experimental data demonstrated that compared to other models,GCA-YOLOV5 exhibits superior accuracy,fewer parameters,lower computational cost,and smaller size.On the embedded device side,an inference performance comparison analysis was conducted between GCA-YOLOV5 and YOLOV5,and the experimental data showed that GCA-YOLOV5 requires fewer computing resources and has stronger parallel task processing capabilities.Therefore,the proposed GCA-YOLOV5 algorithm in this study provides a good approach to lightweight modeling in deep learning-based pest and disease recognition in agriculture.