Research And Application Of Strawberry Fruit Powdery Mildew Recognition Based On Transfer Learning

Strawberry occupies an important position in fruit production and is known as the "Queen of Fruits".However,strawberries are very vulnerable to pests and diseases,and the occurrence of powdery mildew in strawberry fruits will seriously affect the yield and quality of strawberries.Traditional image recognition methods are mostly artificially selecting and extracting disease characteristics from diseased leaves and fruits,which is time-consuming and laborious,and is easy to misjudge due to subjective factors during periods when the characteristics are not obvious.Therefore,this research aims to develop an efficient and accurate disease identification method for strawberry fruit powdery mildew.Convolutional neural networks have strong autonomous feature learning and expression capabilities,and can effectively avoid complex artificial feature extraction processes.In recent years,they have been widely used in plant disease identification.However,most of the current researches focus on rice,corn and other crops,and there is no deep network model for strawberry fruit powdery mildew.Therefore,this research design selects 4 deep network models(VGG-16,Res Net-50,Inception-V3 and Dense Net-121).Using the method of parameter migration,multiple training is performed on the large data set Image Net,and the model training parameters are used in the strawberry fruit powdery mildew disease identification model.The main work and results of this research are as follows:(1)Establishment of image database of strawberry fruit powdery mildew disease.Two images of healthy strawberry fruit and powdery mildew fruit were collected in the field environment,and a strawberry fruit powdery mildew disease image database containing 6,364 images was established.In order to explore whether the redundant information of the field environment affects the feature extraction of the convolutional neural network,this study uses the threshold segmentation based on HSV and the Grab Cut algorithm to achieve the target extraction of strawberry fruit images.In addition,data expansion is achieved by performing affine transformation on image samples,which improves the robustness and accuracy of the network model.(2)Construction and optimization of disease recognition model based on transfer learning.This study proposes 4 improved strawberry fruit powdery mildew disease identification models,based on the migration learning method of parameter migration,reinitializes the last layer of the 4 network models(replaces the Soft Max classification of the original network with the Soft Max classification layer of 2 labels The other layers directly use the weight parameters pre-trained on the large image data set Image Net,and then use the strawberry fruit image database to fine-tune the network parameters.Considering the complexity of the experiment,a total of 36 sets of combined experiments were carried out by combining 3 image types and 3 resolution sizes.Experiments show that the improved Dense Net-121 model based on migration learning has the highest recognition accuracy on thestrawberry fruit powdery mildew test set when the target is extracted and the resolution is512×512,reaching 98.12%.(3)Design and development of strawberry fruit powdery mildew disease identification system.A disease identification system was built based on the Django framework to realize the disease identification of strawberry fruit powdery mildew.The system integrates image input,recognition processing,result display and other functions,providing fruit farmers with an accurate way to identify strawberry fruit powdery mildew disease.The method and application of strawberry fruit powdery mildew disease identification based on migration learning proposed in this paper can achieve effective and accurate identification of powdery mildew,and has certain reference value in the field of crop disease identification.