Research On The Identification Methods Of Tea Leaf Disease Based On Image Characteristics

Tea leaf was one of the most main drinks in china,which was one of the country that the tea planting scale was largest,while the various diseases have been seriously affecting the yield and quality of tea.Not only the traditional methods of disease prevention and treatment were time and labor consuming,but also the governance effects were poor.With the continuous development of computer technology,image processing and pattern recognition in recent years,the automatic diagnosis of plant disease through image characteristics became feasible.In order to achieve the accurate identification of tea leaf disease in the natural environment,three diseases which were common in tea leaf was researched in this paper,they were the tea white scab,tea brown leaf spot and Colletotrichum camelliae respectively.The classification and recognition of leaf diseases based on image characteristics was the key of this passage.Main contents and conclusions are as follows:(1)The complexity of the background of pathological tea leaf images which was gathered under the natural condition and the differences between the feature of three kind of disease spot contribute to the designing of different preprocessing methods for three kinds of disease respectively in this paper.When extracting the disease spot on the tea leaves which were infected with the tea white scab disease,this paper used the method of adaptive filtration to filter the image in the first,gray level transformation,opening function reconstruction,hat operation and Otsu image segmentation and other methods was run and the final effective extraction rate was 100%.When extracting the disease spot on the tea leaves which were infected tea brown leaf spot,this paper used adaptive filtration in color disease images to filter the image firstly,secondly color space transform was conducted,this paper applied the image segmentation method of Otsu to H channel in the next step,eventually through morphology operations such as opening and closing operation,the effective extraction rate was up to 98%.When it comes to the disease spot extracting of Colletotrichum camelliae,the method of adaptive filtration and color space transform in color disease image were also applied in the first step,then after the segmentation method of Otsu to H-S and R channel and the AND operation between this two images,this paper achieved the eventual effective extraction rate of 93%through morphology operations such as opening and closing operation.(2)The feature extraction method was researched after the preprocessing process applied to tea disease spot images as described in(1).The characteristics of RGB and HSV color space was investigated to full use.this paper calculated 1 to 3 moments under six color channel with a total of 18 components of color feature.Gray level co-occurrence matrix was adopted to calculate energy,contrast,correlation,entropy and smooth degree of the disease spot image with five components of texture feature.The area and perimeter of disease spot was helpful to calculate the value of circularity,rectangle degree,complexity and the eccentricity ratio.This paper discussed emphatically on the calculation process of seven moments which was stable under various calculation such as rotation,translation and so on.Finally this paper got a feature set consisted with 34 feature component.(3)Through the use of the 34 extracted characteristics on color,texture and shape,this paper built a 34-12-3 BP network with three layers.After the training and testing process,the network finally attain the recognition rate of three tea leaf diseases:tea white scab,tea brown leaf spot and Colletotrichum camelliae,they were 95.875%,96.625%and 96.000%respectively,and the average recognition rate 96.167%was achieved.This paper adopts random forests to generate 100 decision trees to classify three kinds of diseases,the recognition rates were 98.500%,94.000%and 97.875%,the average recognition rate was 96.792%.LibSVM which was generally used to build support vector machine(SVM)was builded and trained.The test process applied to samples displayed the final recognition rates of 99.000%,90.250%and 94.750%,the average recognition rate was 94.667%.By comparing the recognition rate of three kinds of algorithms,the fact that random forests for three kinds of disease has the highest average recognition rate was obvious.While the recognition rate attained by the random forests method of tea brown leaf spot was relatively low,the stability of the network was slightly worse than the BP neural network;The average recognition rate of BP neural network was slightly lower than random forests,but its recognition rate of all three kinds of diseases can be over 95%,so it's stability was better;For support vector machine(SVM),except the recognition rate of tea white scab was highest,the other aspects were not as good as the first two algorithms.The research content and conclusion of this paper provided a new method for the identification of tea leaf disease in natural condition,and also had a certain reference significance for the identification of other plants disease.