Research On Crop Image Segmentation And Pest Recognition Based On Deep Learning

As China’s primary industry,agriculture occupies an important proportion in China’s economic system.Affected by factors such as abnormal climate and changeable human activities,the outbreak rate of crop diseases and insect pests is high,which seriously affects yields.The traditional identification of crop diseases and pests is still in the manual inspection stage,which is not only time-consuming and labor-intensive,but also extremely inefficient.At present,the identification of crop leaves is one of the effective methods of pest control.Traditional recognition methods are difficult to capture high-level semantic features when the background environment is more complex,and the learning efficiency is low and the recognition effect is poor.The pest recognition method based on deep learning can extract the deep abstract features of the image,and the robustness and noise resistance of the network are strong.However,due to the wide variety of crops and the complex growing environment,it is difficult to improve the accuracy of relying solely on images for identification.Therefore,this thesis combines crop images with their growth environment information,and proposes crop image segmentation and pest recognition methods based on deep learning,which improves the accuracy of crop disease and pest identification and makes the control work more targeted.The main contents and innovations of this article are as follows:(1)In order to improve the accuracy of leaf edge segmentation,this thesis improved the U-Net network and proposed a leaf segmentation algorithm based on AAU-Net(A-Inception+AM-SE+U-Net)network.Firstly,the original convolution is replaced with a-Inception asymmetric multi-channel convolution to achieve the purpose of enlarging the receptive field and extracting edge features at multiple scales.Then,the attention mechanism SE module is introduced in the jump connection part to enrich the feature diversity and reduce the noise caused by upsampling,so that the network can better combine the image high-level and low-level semantics.Experimental results show that the MIo U value of the proposed segmentation algorithm reaches 74.8% on open data sets,which is highly competitive.On private data sets,the segmentation accuracy of this model is 91.2%,which is better than that of U-Net network.(2)In order to improve the accuracy of pest identification,this thesis proposes a pest identification model based on multi-mode fusion.Firstly,the global feature of blade image was extracted by CNN-non-local.Then,DNN is used to extract features from sensor environment data.Finally,the classification of crop diseases and insect pests is identified by deep neural network after the bi-modal feature fusion of the two feature vectors.Experiments show that the accuracy of the recognition model based on multi-mode fusion is up to 91.3%,and the recognition effect is better than that of the model based on single mode input.Generalization experiments show that the recognition algorithm in this thesis is portable.(3)Based on the above research content,this thesis designed and developed a tomato disease and insect pest identification system,which realized real-time monitoring of images and sensors and automatic identification of diseases and insect pests,and improved the popularization process of intelligent agriculture.Through testing the accuracy and function of the system,the result shows that the system can achieve the expected effect.This thesis includes 44 figures,17 tables,and 80 references.