A Research On Semantic Annotation Of Agricultural Products Images Based On Deep Learning

The field of computer vision and natural language processing are the hot research directions nowadays,while the field of image semantic annotation combines these two different directions to express image information in the form of text language.Rice is one of the most important food crops in China and a hot issue in agricultural economy.Nowadays,on the one hand,automatic image semantic annotation has limitations in application scenarios,and most experts and scholars use open data sets to conduct research in the field of life scenes.On the other hand,the diagnosis of rice plant diseases and insect pests mainly depends on the subjective experience of farmers.Therefore,in this study,computer vision and natural language processing were used to realize the image recognition of rice pests and diseases and the description of pathological features,so that a more efficient and accurate diagnosis method could effectively reduce the economic losses caused by rice disease and yield reduction to a certain extent.The main contents of this thesis include the following three aspects :(1)10 common rice disease and insect pest images were collected and preprocessed to build a Rice＿2k dataset for model training,including 2238 samples and 11190 text sentence annotation sets for semantic annotation training;(2)based on convolution model of neural network model of plant diseases and insect pests of rice image classification,vector by adjusting the maximum number of training,the parameters,the selected image of a training sample,the classification of the activation function to choose the best effect of parameter values,with the classification effect after parameter adjustment of the BP neural network model,It was found that the convolutional neural network model with parameter adjustment and experimental comparison had better effect on rice image classification of pests and diseases.(3)The automatic semantic annotation model of rice disease and insect pest images based on Show Attend and Tell was established.The annotation effects were compared with different image semantic annotation models and different data sets.The results showed that: The Show Attend and Tell model has good annotation accuracy on the Rice＿2k dataset created in this thesis.