Research On Maize Disease Recognition Based On Deep Learning

Corn is an important food and feed crop,and it is also the crop with the highest yields in the world.Its planting area is second only to rice and wheat.The corn with excellent environmental adaptability is more drought-tolerant,cold-tolerant,and barren tolerant than other food crops such as rice and wheat.Since the beginning of the 21st century,the population has been increasing,improving the quality of agricultural products and increasing yield have become the key to solve the food shortage.However,corn diseases severely restrict the production of corn.Therefore,it has become one of the current essential tasks to prevent and control corn diseases and reduce the losses caused by the diseases on corn production.In recent years,deep learning has developed rapidly.Compared with traditional image recognition algorithms that rely on artificially designed feature extraction,deep learning methods can automatically extract features,have good scalability,and have broad application prospects in the agricultural field.In this paper,based on the deep convolutional neural network algorithm,the identification of corn leaf diseases was carried out.Mainly includes the following work:(1)Use data enhancement technology to enrich the number of images and improve the recognition effect.(2)Construct channel attention module and spatial attention module according to attention thought.The channel attention module realizes the constraint enhancement of each feature channel by distinguishing the importance of different feature channels.The spatial attention module extracts the spatial position information of the feature to constrain and enhance the feature in the spatial position.In turn,it enhances the network’s attention to important features and suppresses the network’s attention to non-important features.After experimental comparison,the parameter amount of the finally constructed network model is 28.08M,and the average recognition rate is 96.34%.(3)Combined with residual thoughts,construct a DARNet model with attention thoughts and residual thoughts.With the help of residual thinking,the feature value reduction caused by the attention module can be avoided,the degradation of the network can be avoided,and so on.Taking corn leaf images with spot,small spot,rust,gray spot,southern rust and healthy respectively as the research objects,the average recognition rate of DARNet model is 1.16%to 9.38%higher than that of VGG16,Resnet50,ShuffleNet,SqueezeNet and SeNet models,and the recognition rate of various diseases is 0.22%～1 8.92%higher.(4)Divide each type of disease into general and severe,and study the recognition effect of DARNet model on the severity of the disease.The prediction results given by the DARNet model include the type of disease and the severity of the disease.From the perspective of identification results,the reliability of the disease severity(general or serious)in the prediction results are on average 84.42%and 90.58%.If the prediction results in the disease category are not corn leaf spot,then the reliability of the disease category in the prediction results is greater than 95%(91.69%for corn leaf spot).(5)Using Flask to develop Web applications,with functions such as registration,login,identification,etc.,to realize the identification of corn leaf diseases on multiple platforms.Under Tesla P4,only one image is recognized at a time.After testing,it takes 7.89s to continuously recognize 100 images,and it takes 79ms to recognize each image on average.Use Docker to encapsulate the relevant code,environment,etc.,reducing the difficulty of deployment.