Research On New Semantic Segmentation Model And Weed Recognition Based On Deep Learnin

Corn is an important food product of the country,is severely damaged in agricultural production.Under extreme conditions,corn will reduce production by more than 20% because of weeds.The traditional method of applying large-area total spraying pesticide herbicides can not apply weeds based on demand,which is easily causing a large amount of pesticide waste,high cost of production,severe pollution of the environment and other problems,thus affecting the yield and quality of corn.And the existing mechanical weeding,chemical weeding and other methods have a variety of shortcomings,such as environmental pollution,increase labor costs and so on,so efficient intelligent weeding technology research is very practical significance.In order to solve the above problems,precise variable application of maize weeds is needed.Identification of weeds in crops by image recognition technology and removal of weeds by mechanical or chemical agents have the potential to effectively reduce weed damage to crops and improve crop yield and quality.At the same time,in comparison to the ordinary continuous spray,the use of image recognition technology for weed control can also reduce environmental pollution,which is conducive to sustainable agricultural development.Based on the above problems,the objective of this paper is to perform rapid and accurate identification of maize seedlings and weeds,and utilizes computer vision and deep learning technology to analyse and investigate the method that can automatically identify maize seedlings and weeds.Specific research contents are as follows:(1)Analysis and comparison of weed and maize emergence rate and requirements of subsequent processing on image quality,so as to determine the image selection.Set the time and the height and Angle of the shot during the acquisition process;Using geometric transformation:image flipping,image scaling,image normalization;Color transform and gamma transform enhance the image data.Label Me software was used to manually label the images,and corn and weeds were distinguished by different colors.(2)In order to improve the identification and localization accuracy of maize and weeds and densely distributed weeds in the field under complex environment,a modification was proposed Weed segmentation algorithm into Res Net50 network.First of all,in the process of feature extraction,the image is made to conform to the corresponding display area by sampling under the convolution residual layer.After a transposition convolution step of up-sampling,the image size is doubled each time,and finally the image size is restored to the original size,which solves the accuracy problem.Finally,the Softmax layer is used for output and the classification diagram is obtained.(3)In order to study the effect of different modules on weed segmentation and recognition models with feature fusion,different trunk extraction was selected first.The modules are VGG network and Res Net network.By replacing the original Unet extraction network with Res Net network,the model has a better extraction ability.On the basis of Res Net as the extraction module of Unet segmentation network,the CA attention module and FFM feature fusion module are respectively ablation experiments.Then,CA attention module was used to further strengthen the extraction network.Finally,FFM feature fusion module was used to fuse the feature information of different semantics.It was proved that FFM feature fusion module and CA attention mechanism could improve the segmentation effect of maize and weeds.