Recognition Technology And Application Of Weed In Corn Field Based On Deep Learning

Our country is a big country with abundant agricultural products.Attaching importance to the development of agriculture is the foundation of people's security and the key to governing the country.The development of precision agriculture and the realization of automation and informationization of agricultural development has become an inevitable trend of agricultural development.Relying on image recognition technology to identify field weeds in crops,and then using mechanical weeding or chemical weeding to remove weeds can effectively reduce the damage of field weeds to crops and improve the yield and quality of crops.At the same time,compared with the traditional continuous spraying method,weeding with image recognition technology can also reduce environmental pollution and promote the sustainable development of agriculture.Corn is an important food crop in China.In our country's food crops,the planting area of corn is third only to rice and wheat.In 2018,the planting area of corn reached 632 million mu(equivalent to 42 million hectares).In order to improve the accuracy of weed image recognition,this paper takes corn and its associated weeds as the research object,combines the strong feature extraction ability of deep convolution network and the characteristics of Hash code for easy storage and fast retrieval.The main research contents of this paper are as follows:(1)A weed classification and recognition network is proposed.This network combines the strong feature extraction ability of deep convolution network and the characteristics of hash code for easy storage and fast retrieval,effectively compresses high-dimensional weed features.After the full connection layer,a hash layer structure is designed and added to transform the eigenvalues of the full connection layer of the extracted maize and weed images.Then the spatial distance between the corresponding images is calculated and sorted in descending order to realize the classification of corn and its associated weeds.The structure and Super-parameters of weed classification and recognition network were determined by designing relevant comparative experiments.Experiments show that the average accuracy of this method is 97.73% and the correct recognition rates of corn,cerer,huicai,suocao and zaoshuhe are 98.67%,98%,96.67%,98% and 97.33% respectively.It shows that good classification results are obtained.(2)A new weed segmentation network based on Mask R-CNN is implemented.The common image segmentation network and its structure and training process are studied.Using the mAP(mean average precision)index to test the weed segmentation network based on Mask R-CNN,the experimental results show that when IoU = 0.5,the method can achieve 0.8133 mAP value.The experimental results also show that the edge segmentation of various plant images is accurate,the effect is remarkable,and the expected goal is achieved.(3)A weed segmentation network based on Mask R-CNN and hash code is proposed.Combining the above two networks,the part of classification network of Mask R-CNN is replaced by the first network.Initializing the classification network and segmentation network by higher precision model's parameters obtained on COCO dataset and corn and it's associated weeds dataset.Then fine-tuning the model on the data set of maize and its associated weeds.Finally,it was tested in corn field.The results showed that the correct recognition rate of weed and the wrong recognition rate of corn are 92.06% and 1.4% respectively in the application.It shows that good results were obtained by using the method in this paper.