Research On Improved Lightweight Cassava Disease Identification Based On YOLOV4

Cassava is an important cash crop and food crop widely planted in southern China.In recent years,the production of cassava cultivation in China has gradually decreased.In addition to being influenced by factors such as cassava prices,the prevention and control of cassava diseases is also a major issue affecting cassava production and the cassava planting industry.Traditional cassava disease prevention and control mainly rely on manual detection,which has the drawbacks of low detection efficiency,untimely detection,and high labor costs.Compared to traditional methods,deep learning based crop disease detection is more stable and efficient,but it also has the disadvantages of high computational and parameter requirements for most neural networks,high hardware and software configuration requirements for recognition equipment,and large model memory,making it difficult to apply in practical scenarios.To address the above issues,reduce the model memory,reduce the required software and hardware configuration,and use it for cassava disease detection.This study proposes a KM+YOLOv4 lightweight model that is optimized and improved through multi-scale feature fusion and anchor box algorithm.The main work of this paper is as follows:1.Collect cassava leaf diseases in dense scenarios and create a dataset of dense cassava diseases.Optimize the Kmeans anchor box algorithm from randomly selecting cluster points to selecting the farthest point from the previous cluster point as the new cluster point.Further optimize the distance calculation formula of the clustering algorithm,optimize the distance from direct calculation to Euclidean distance,optimize the clustering algorithm to the improved Kmeans++algorithm,and reconstruct anchor boxes suitable for cassava diseases to improve the detection accuracy of the model.2.Improve the YOLOv4 model backbone network by replacing the CSPPark Net53 ordinary convolutional backbone network with the Mobile Netv3 deep separable convolutional lightweight network as the new MYOLOv4 lightweight model.Research has found that although MYOLOv4 lightweight network can reduce model memory,it has a higher accuracy loss problem than ordinary convolution in dense detection scenarios.To compensate for accuracy loss,add 5 × The deep separable convolution of 5 is fused with the MYOLOv4 backbone network to form an improved lightweight multi-scale feature fusion network,and combined with the improved Kmeans++algorithm to form the KM+YOLOv4 model.3.Apply the dense cassava disease dataset produced to the KM+YOLOv4model.After learning and training,the KM+YOLOv4 model has a memory of64 MB,which is reduced by 181 MB compared to the original YOLOv4 model’s memory of 245 MB.The calculation parameters and model size of the reduced backbone network are about 75% of the original YOLOv4,and the detection performance is improved by about 1.4% compared to the lightweight model MYOLOv4.This model can be better applied to mobile and embedded systems,helping to identify cassava diseases,prevent and control cassava diseases,and improve control efficiency.