Research And Implementation Of Recognition And Location Of Strawberry Leaf Powdery Mildew

The strawberry market in China is vast,and the strawberry cultivation bases across the country are increasing year by year.However,strawberries are extremely susceptible to disease attack during cultivation,resulting in large-scale production cuts.Strawberry powdery mildew is very harmful to the strawberry industry,especially in Northeast China,where strawberry is more susceptible to powdery mildew.Because strawberry powdery mildew mainly harms the leaves of strawberry and leaves are important organs to ensure the healthy growth of strawberry fruit,the identification of powdery mildew in strawberry leaves is the cornerstone of controlling powdery mildew and the key to guarantee the yield of strawberry.Therefore,this study aims to design and implement a model that can quickly and accurately identify powdery mildew and locate lesions in strawberry leaves.Since witnessing Alex Net’s superior classification capabilities in 2012,various industries have begun to conduct intensive research on convolutional neural networks.Agricultural researchers around the world are also beginning to focus on convolutional neural networks.Although the research results of the combination of agriculture and convolutional neural networks have sprung up,there are few studies combining the strawberry industry with the convolutional neural network.In view of the remarkable achievements of convolutional neural networks in the field of computer vision and the importance of strawberry powdery mildew recognition and lesion location in the strawberry industry,this study uses the identification of powdery mildew and the location of lesions in strawberry leaves as research purposes.The strawberry leaf image is the research object,and the convolutional neural network is used as the research method.The main work and results of this research are as follows:(1)At the Strawberry Research Base of Shenyang Agricultural University,10,000 images of strawberry healthy leaves and 8000 images of strawberry powdery mildew were collected to establish strawberry leaf dataset.Both healthy leaf images and powdery mildew images were collected under natural light in a strawberry greenhouse.(2)Considering that the strawberry leaf dataset sample size is small and the characteristics of the powdery mildew leaf are not easy to learn,migration learning is used for the disease recognition model.A comparative analysis of several Image Net-trained network models for feature extraction of powdery mildew leaves.After comprehensively considering the correct rate and calculation speed,it was decided to use Dense Net121-No Top as the feature extractor.Six different types of classifiers were designed.After testing and comparative analysis,GAPSoftmax was found to have the best classification effect.Finally,a strawberry leaf powdery mildew recognition model with Dense Net121-No Top as the feature extractor and GAPSoftmax as the classifier is proposed,which is abbreviated as DA network model.After experiment,the correct recognition rate of strawberry leaf powdery mildew by DA network model reached 98.9%.(3)Considering that the samples of the strawberry leaf dataset are not marked with lesions,only weak supervised learning can be used to achieve lesion localization.After comparing and analyzing the existing weak supervised learning methods,it was decided to adopt the GAIN network model.The GAIN network model is combined with the DA network model to obtain a DAG network model.The DAG network model can train both recognition and positioning capabilities.Tests have proved that the DAG network model can roughly define the area of powdery mildew lesions.(4)Develop Android APP and Web pages and create a RESTful API on the Ubuntu system that provides a unified interface for different platform clients,ultimately implementing the strawberry leaf powdery mildew diagnostic system.