Research On Pear Leaf Disease Recognition Based On Deep Learning

Deep learning is a branch of artificial intelligence.In recent years,with its advantages of automatic learning and feature extraction,it has been widely concerned by academic and industrial circles.It has been widely used in image and video processing,voice processing and natural language processing.At the same time,it has also become a research hotspot in the field of agricultural plant protection,such as plant disease recognition and pest range assessment,etc.However,there are still some problems in the field of plant disease identification using deep learning,such as incomplete plant disease dataset,insufficient research on factors affecting disease recognition performance of deep learning network model and so on,which will restrict the promotion and application of deep learning in plant disease prevention and control practice.In order to study and solve the above problems,this paper takes pear leaf disease as an example,starting from the two dimensions of constructing disease dataset and exploring various factors that affect the performance of deep learning network model to identify disease,using deep learning to carry out related research.The main research work and conclusions are as follows:(1)Pear Leaf Disease Dataset was constructed,and the effect of disease dataset on disease recognition accuracy of deep learning network model was studied and discussed.Aiming at the three common Septoria piricola,Alternaria alternate and Gymnosporangium haracannum in Shanxi Province,which are mainly single diseases,7620 effective pear leaf samples were collected from indoor environment,outdoor field,front and back of leaf.According to different experimental objectives,a variety of different types of Pear Leaf Disease Datasets were constructed,such as PLDD,PDD2018,PDIRE,etc.Taking PDD2018 as a typical example,it has the following five significant characteristics: 1)diverse collection environment;2)strong regional characteristics;3)rich varieties;4)different stages of disease;5)paying attention to the back of pear leaf.As far as we know,we have not found any plant leaf disease dataset with similar characteristics of PDD2018.Through a large number of experiments,this paper studies and discusses the influence of disease dataset capacity,disease sample image characteristics and disease sample collection environment on disease recognition accuracy of deep learning network model.The obtained results show that the capacity of disease dataset is an important factor affecting the accuracy of disease recognition in deep learning network model.Using transfer learning to study plant disease recognition,taking ResNet50 as an example,when the number of disease samples of each kind is maintained at more than 350,ResNet50 can achieve higher disease recognition accuracy,which provides empirical support for the plant leaf disease sample collection quantity requirements.The obtained results show that the image characteristics of five disease samples that may reduce the disease recognition accuracy of deep learning network model are as follows: 1)the leaf area ratio is too small;2)the color feature similarity of different diseases;3)the leaf is obviously damaged;4)the number of lesion on leaf is very small or the lesion area is positively small;5)there are attachments or dehydration on the leaf surface.It has important reference significance for constructing high quality plant leaf datasets.The obtained results show that the image background(acquisition environment)of disease samples directly affects the disease recognition accuracy of deep learning network model.When the collection environment or image background of disease samples changes,the disease recognition accuracy will decrease significantly when the model which has not been trained based on disease datasets of similar environment is used to identify such disease samples.Taking Res Net50 as an example,the sensitivity of Septoria piricola was the highest,followed by Alternaria alternate and Gymnosporangium haracannum.(2)A technical approach of "deep learning network model + input image resolution" suitable for pear leaf disease recognition was proposed.In this paper,through a large number of experiments and comparative studies,under the condition of the same hardware resources,"ResNet50 + 600 × 600" training time is moderate,and the highest disease recognition accuracy rate is 98.7%,the disease recognition accuracy of Septoria piricola,Alternaria alternate and Gymnosporangium haracannum are 99.44%,98.43% and 97.67%,respectively,which is the most suitable for pear leaf disease recognition.The "ResNet50 + 600 × 600" makes it possible to build a deep learning disease recognition model based on small dataset by modifying the parameters of deep learning network model by using transfer learning.It can directly take pear leaf sample image as the input of deep learning network model,and provide a technical means,which is helpful for the rapid application and promotion of deep learning technology in the practice of disease prevention and control.(3)Various factors affecting the performance of deep learning network model to identify pear leaf diseases were studied and discussed.This paper studies and discusses the impact of disease sample image resolution and model parameter settings on performance indicators such as the disease recognition accuracy of deep learning network models: It is pointed out that the input disease image resolution is obviously positively correlated with the disease recognition accuracy.Too low resolution can easily lead to over-fitting problems in the model,and too high resolution can easily lead to memory overflow problems.Modifying the parameters of the deep learning network model and increasing the input disease image resolution could improve the disease recognition accuracy.And,there is an obvious positive correlation between the input disease image resolution and training time-consuming;it is also pointed out that the parameters of the deep learning network model,such as the number of training rounds and the number of network layers,do not necessarily have a positive correlation with its disease recognition accuracy.This paper studies and explores the influence of pear leaf front and back disease spots on the disease recognition accuracy of deep learning network model for the first time.The evaluation criterion of disease recognition consistency(DRC)and disease recognition correct consistency(DRCC)are proposed to evaluate the consistency of disease identification results of the two models.The experimental results show that the characteristics of the front and back of pear leaf lesions play the same role in disease recognition,so,the pear disease recognition model can be established based on the front of leaf or the back.The combination of the front and back is more conducive to pear leaf disease recognition.Finally,the identification characteristics of different pear leaf diseases were studied and discussed.