Research On Tea Disease Image Recognition In Natural Environmen

At present,most of the tea disease identification research based on machine learning and deep learning is aimed at leaf images in simple background,while tea disease images in natural environment not only have complex backgrounds,but also uneven illumination,and the color,size and shape of disease spots at different growth stages are also different,which greatly affects the identification rate and detection efficiency of tea diseases.This article takes the image of tea diseases in the natural environment as the research object.Based on different deep learning and machine learning algorithms carry out identification and testing research.It can provide technical support for the intelligent identification and precise application of tea diseases,and the main research contents and conclusions are as follows:(1)In order to improve the detection accuracy and efficiency of tea disease spots in natural environment,this paper proposes a tea disease spot detection method based on improved YOLOv5 s.The early tea disease is not obvious and the spot size is small,and the downsampling operation of the YOLOv5 s backbone feature extraction network makes the shallow feature information of the early disease lost,which is easy to lead to the missed detection of early disease spots.In view of the above situation,the CA attention mechanism is introduced into the YOLOv5 s network structure,and by embedding the position information into the channel attention,the model is told what content and location needs to pay more attention to,and the feature information can be effectively extracted for small targets and dense targets to improve the target detection rate.A weighted bidirectional feature pyramid network(Bi FPN)that can learn weights and learn the importance of different input features is introduced to achieve higher-level feature fusion,and a small target detection head is added to the detection layer,which improves the robustness of small target and overlapping target recognition.The experimental results show that the m PA of the proposed method for the detection of different diseases reaches 97.3%,the detection accuracy of early diseases is improved by0.9%,and the detection effect of tea disease spots under different weather conditions is also good.(2)In order to remove the complex background of diseased leaves in the natural environment,the segmentation performance of U-Net,DeepLabV3+,PSPNet and HRNet on tea disease leaves was compared and analyzed,and the network suitable for disease leaf segmentation was optimized,and tea disease leaf segmentation experiments were carried out under different hyperparameter combination conditions,and the results showed that DeepLabV3+ network had the best segmentation performance in tea disease leaf images,and the average cross-union ratio of diseased leaves reached 97.82%.Then,machine learning methods(support vector machine,random forest)are used to identify tea disease images after removing complex backgrounds.By extracting different color and texture features of disease images for comparative experiments,the effects of eight different feature extraction methods(LBP,HOG,GLCM,COLOR,LBP+GLCM,LBP+COLOR,COLOR+GLCM,ALL)on the disease recognition effect were compared and analyzed,and it was found that the optimal recognition rate of different diseases was related to the feature extraction method.Among the two machine learning algorithms,the COLOR+GLCM+SVM algorithm has the best recognition effect for different diseases,with an accuracy rate of 90.39%.However,the accuracy of machine learning algorithms for similar diseases is not high,which is easy to cause misidentification.(3)The recognition performance of five lightweight deep learning networks MobileNetV2,ShuffleNetv2,EfficientNetv2,RegNet and DenseNet on tea disease images in the natural environment was compared and analyzed,and the ShuffleNetV2 model had the best effect by considering the accuracy,parameter quantity and calculation amount,with an accuracy rate of 95.7%,a parameter quantity of 1.26 M,and FLOPs of 303.38 on the test set.Transfer learning was carried out on ShuffleNetV2 network,and the problem of high false identification rate of similar diseases was improved by fine-tuning the model parameters,and the recognition accuracy in the test set reached 99.8%,which was 4.1% higher than that of the model before transfer learning.