Identification Of Quinoa Leaf Diseases Based On Image Processing

Quinoa leaf spot and downy mildew are two diseases with wide influence range in quinoa diseases,and the propagation speed is gradually accelerating.If quinoa leaf spot and downy mildew are not found in time,quinoa production will suffer great losses or even harvest.Nowadays,the detection of quinoa diseases is mainly conducted by professionals or farmers.This method not only increases the labor cost and consumes a lot of time,but also causes identification errors due to human subjective factors,which cannot meet the requirements of modern agriculture.Therefore,it is of great significance to provide convenient and effective methods for quinoa leaf disease identification and improve the accuracy of quinoa disease identification.Based on the images of quinoa leaves,this paper proposes a new method for quinoa disease identification,which improves the identification accuracy of quinoa leaves.Firstly,through the study of color space technology,leaf segmentation method and morphological processing method,the leaf segmentation model of quinoa disease was established,and the segmentation of quinoa leaf in complex background was realized.Then,through the feature extraction of quinoa leaves and the image processing method and machine learning method of quinoa diseases,the identification model of quinoa diseases was established to identify healthy quinoa,quinoa leaf spot and quinoa downy mildew.The main results of the study are as follows :(1)In order to solve the background separation problem of quinoa plants in complex environment,the leaf segmentation method of quinoa was studied.The modified super-green feature algorithm was used to separate quinoa plants from complex backgrounds.The stem in quinoa image is eliminated after ultra-green feature processing by corrosion treatment,and the gap left after corrosion operation is filled by expansion treatment.The expanded image was compared with the original image to transform color,which made up for the shortcomings of fuzzy edge and obtained a clearer image of quinoa leaves.The quinoa leaf image was binarized,and the quinoa leaf was segmented using the breadth search algorithm to extract the single quinoa leaf,and the quinoa leaf information was completely retained.Finally,the quinoa leaf data set was expanded.(2)In order to better obtain leaf information of quinoa,the color moment and gray level co-occurrence matrix method was used to extract feature parameters based on leaf color and texture features.The eigenvalues of R,G and B channels in quinoa leaf image are selected by color moment,and two effective color features are extracted.The changes of entropy,energy,inertia moment and correlation parameters of quinoa leaf images in different directions were verified by gray level co-occurrence matrix method.Experiments show that the eigenvalues are not disturbed by the change of image direction;the parameters of four texture feature values at different gray levels are further analyzed,and the gray level with less hardware burden is selected under the condition of minimizing information loss.Finally,six effective feature parameters are extracted for the input of the subsequent recognition model.(3)In order to improve the identification accuracy of quinoa diseases,three quinoa leaf disease identification models were constructed.Taking normal quinoa leaf image,quinoa leaf spot image and quinoa downy mildew image as training samples,three recognition algorithms are compared and studied.The average recognition accuracy of support vector machine algorithm is 93.43 %,and the recognition accuracy is the highest.The algorithm is very suitable for the classification calculation of a small number of samples,and has a certain dependence on equipment.In view of the structural characteristics of BP neural network,this algorithm has obvious advantages in large sample recognition.Under the condition of sufficient network training,the recognition of single state can achieve the recognition results of small samples,and the overall average recognition accuracy can reach 89.03 %.Convolutional neural network has less requirements for the initial image.Its unique weight sharing and down-sampling methods have high invariance to different image sizes,angles and feature positions,and can achieve good accuracy in a short time.The overall average recognition accuracy can reach 91.53 %.It can be seen that the three recognition algorithms are feasible and efficient for leaf recognition of quinoa in different conditions...