Identification And Classification Of Potato Late Blight Leaf Diseases Based On Deep Learnin

Late blight is the most common disease affecting potato plants,which severely impacts potato yields.Late blight spreads quickly when the weather is humid,and if not treated in a timely manner,the disease can rapidly spread and cause entire plants to wither and die,resulting in significant economic losses.Therefore,accurately identifying the level of late blight is a crucial prerequisite for timely and effective prevention and control,and is an important means of ensuring potato yields.Traditional methods of identification,such as manual identification,instrument identification,and machine learning-based identification models,are costly,inefficient,and cumbersome,making it urgent to design an economical,efficient,and simple identification model to replace traditional identification methods.This study converts the problem of identifying and grading late blight on potato leaves into image classification and semantic segmentation tasks in computer vision.Deep learning techniques are used to design potato leaf classification models and semantic segmentation models,improving the efficiency and accuracy of disease level identification.The overall scheme of the study consists of three stages: dataset preprocessing,fine-grained disease recognition,and lesion segmentation and disease grading.In the dataset preprocessing stage,a mixed dataset is created to enhance the model’s robustness for practical applications.To address the problem of missed and erroneous annotations due to low-resolution images,a super-resolution image reconstruction algorithm is designed by combining chameleon extraordinary visual function and bidirectional recurrent neural networks.Experimental results show that the proposed algorithm performs better than other mainstream algorithms in terms of peak signal-to-noise ratio,structural similarity evaluation index,and subjective quality score.The algorithm is used to reconstruct high-quality annotated samples from low-resolution images for the following two stages.In the fine-grained disease recognition stage,transfer learning is used to address the problem of similar representation and small sample size for early and late blight.The transfer learning results on classic recognition models show that residual networks have the strongest transfer learning ability,with an accuracy of up to 99%.In the lesion segmentation and disease grading stage,Deep Lab V3 model is modified by adding a spatial sensitivity auxiliary channel and modifying the loss function to the sum of two parts: the main channel and the spatial sensitivity auxiliary channel.The ratio of the segmented lesion area to the leaf area is quantified and compared with the grading standards to determine the disease level.Experimental results show that the improved Deep Lab V3 model performs better than other comparative algorithms in terms of lesion edge and small target segmentation,with pixel accuracy and average intersection-over-union evaluation indices of 97.2% and 91.7%,respectively.The accuracy of disease level determination for 18 test samples reaches 94.4%.The deep learning-based potato late blight leaf image recognition and segmentation model can meet the requirements of practical applications with high accuracy,achieving automated disease level determination and providing reliable basis for disease prevention and control.