Research On Diagnosis Method Of Hemp Leaf Stress Under Few Sample Data

Hemp is an important fiber crop with both industrial development value and medicinal development value.During the growth and development of hemp,the s tress of hemp leaves can easily cause the decline of hemp quality and yield,which seriously affects the sustainable and healthy development of China ’s hemp industry.Because of the complexity of the field environment,collecting images of hemp leaf stress is not only tedious and laborious,but the data collected is also limited.The accuracy of using few sample data to train deep learning models to diagnose hemp leaf stress is low and prone to overfitting.Therefore,based on the few sample data from the foliar stress of hemp,a deep diagnostic model is needed to accurately diagnose the type of foliar stress of hemp.In this paper,six types of hemp leaf stress collected in the hemp field were studied,including white scab,brown spot and downy mildew unde r biotic stress as well as potassium deficiency,nitrogen deficiency and water deficiency under abiotic stress.The deep learning technology was used to diagnose the type of hemp leaf stress on the few sample data of hemp leaf stress.The main work is as f ollows:Firstly,aiming at the few sample problem of hemp leaf stress,an improved Siamese Network is proposed to diagnose hemp leaf stress,the learning rate attenuation method is changed and the weight attenuation term is added to the loss function.The model was formed with data about the stress of hemp leaves collected by the Heilongjiang Academy of Agricultural Sciences at various times.The experimental results show that the improved Siamese Network model achieves 90.83% and 86.67% accuracy in coarse-grained and finegrained diagnosis of hemp leaf stress,respectively.Finally,the model achieved a generalization accuracy of 69.17% on the generalization test set collected from Harbin Institute of Industrial Technology in Heilongjiang Province.Secondly,aiming at the few sample and complex background problems of hemp leaf stress,a two-stage method based on improved Yolo X and Siamese Network fusion is proposed to diagnose hemp leaf stress,which replaces the loss function and activation function of Yolo X and increases the post-processing of output.In the first step,the improved Yolo X was used to locate and clip the stress sample block to reduce interference from complicated background information,and its m AP was 91.77%.In the second stage,the improv ed Siamese Network was trained by clipping the stress leaf block data to diagnose the stress type,and the accuracy of coarse-grained and fine-grained diagnosis was 98.11% and 96.21%,respectively.In addition,the influence of background processing method on diagnostic accuracy is compared,and the effectiveness of background clipping method is verified by t-SNE method.Finally,the generalization performance of the model is verified,and the diagnostic accuracy on the generalization test set reaches 75.54 %.Finally,a two-stage method for optimizing the fusion of UNet and Siamese network is proposed to diagnose the stress of hemp leaves,aiming at the problems of few samples,complex background and multiple stresses in one image.The function extraction ne twork and the UNet learning rate mitigation mode are optimised.In the first stage,the optimized UNet was used to locate and process the stress parts of hemp leaves,and the interference of complex background and leaf occlusion was avoided to the greatest extent.The positioning m PA reached 94.59% and m Io U reached 90.26%.In the second stage,the improved Siamese Network was trained by using the data of hemp leaf stress parts.The model achieved 99.37%and 98.74% accuracy in coarse-grained and fine-grained diagnosis of hemp leaf stress,respectively.Finally,the generalization performance of the model is checked,and the diagnostic accuracy on the generalization test set reaches 80.83%.The multi-stage hemp leaf stress diagnosis model proposed in this pape r solves the problem of few sample,complex background and multi-stress of hemp leaves in the field.Aiming at the biotic and abiotic stress of hemp leaves,it provides digital diagnosis and comprehensive prevention and control scheme,which provides new d irection and ideas for the study of hemp leaf stress in China.The model supports the design of applications to provide technical support for lightweight network deployment to smart devices.