Research On Segmentation And Localization Of Crop And Weed Under Low Contrast

Weeds are one of the important factors affecting crop yield and quality.Traditional manual spraying and mechanical weeding have the problems of high labor cost and damage to crop plants,and untargeted whole-area spraying operations have also brought many negative problems such as crop damage,weed resistance and environmental pollution.At present,targeted weeding machinery combined with machine vision can be used for mechanical or chemical control of weeds in the field,and its accuracy lies in whether crops and weeds can be accurately segmented and located.However,the complex and changeable environment in the field often leads to low contrast between the background and the edge of the plant.The traditional visual algorithm is difficult to effectively distinguish the background and the plant in the crop image.In addition,the symbiosis of crops and weeds and the evolution of weed mimicry crops in the field make the phenotypic characteristics of crops and weeds show low contrast.The discriminative information is only hidden in tiny details,which increases the difficulty in obtaining the characteristic difference between crops and weeds to achieve crop/ weed segmentation,thus affecting the accuracy of field targeted weeding.Therefore,this thesis obtains field crop and weed data sets based on field working conditions,and mines essential feature learning methods from image data based on deep learning technology to study crop / weed segmentation and localization methods under low contrast,so as to reduce the influence of complex environment and crop weed morphology on the accuracy of targeted mechanical weeding.The main research contents include:(1)Research on segmentation model of field seedling plant.In order to eliminate the interference of complex background on seedling plant segmentation,aiming at the problem of edge loss in plant semantic segmentation under low contrast conditions,this paper proposes a seedling plant segmentation network model(RECNN)that combines regional semantics and edge information,and combines edge extraction and regional semantic segmentation tasks to make up for missing edge features in segmentation results.RE-CNN uses UNet as the backbone network to extract the semantic information of the plant area,and constructs the edge perception module(EAM)based on the side depth supervision mechanism to guide the backbone network to perceive the edge information of the plant when extracting the regional features.The feature fusion module(FFM)is constructed by using the spatial hole feature pyramid,and the features extracted by the backbone network and the edge perception module are fused.Finally,the network model is optimized by combining edge perception loss and feature fusion loss.The experimental results show that the accuracy of pixel segmentation of RE-CNN at seedling stage is as high as 97.50 %,and the average intersection ratio is 93.20 %.It has good segmentation effect and generalization ability for seedling plants in natural environment,and provides a reliable research object for the next crop weed segmentation and localization.(2)Study on crop and weed segmentation model.In order to reduce the influence of seedling similarity on crop and weed segmentation accuracy,a strip convolutional network model(SC-Net)is proposed in this paper to identify the subtle feature differences between crops and weeds,so as to realize seedling segmentation.The SC-Net consists of a backbone network and a long-jump connection.The backbone network includes parallel multi-scale convolution blocks(PMCB),strip multiscale convolution blocks(SMCB)and strip pooling attention blocks(SPA).Parallel multiscale convolution can extract multi-scale target features.The strip convolution enhances the model ’s perception of the difference in the details of the seedlings by expanding the convolution perception area,while the strip pooling attention module is used to capture the long-distance context semantic information to obtain the strong feature appearance of the seedlings image.By constructing the attention fusion module,the long jump connection fuses the feature maps from the encoder and decoder,screens the effective features,and optimizes the segmentation results.The experimental results show that SC-Net achieves an average intersection-union ratio of 87.50 % and 90.84 % in the field crop weed dataset and the public dataset Boni Rob,respectively,which can provide reliable data support for field targeted weed control.(2)Study on the localization method of stem base center of crop plants.Based on the results of the crop weed segmentation model,the localization method of the stem base center of the crop plant is studied.Taking rice seedlings with various canopy morphology as an example,a localization algorithm based on local area and skeleton structure was designed according to the morphological characteristics of rice plant growth.The background noise in the binary image of crop plants was removed by preprocessing and morphological processing,and the equal interval point and the intersection point of rice plant skeleton in the rice plant area were selected as possible localization points.According to the possible position of the stem base center in the rice plant area,a screening algorithm was designed to calculate the localization coordinates.At the same time,in order to prevent the screened points from gathering away from the center of the stem base due to the deviation of the rice plant leaves to one side,a screening algorithm was designed to optimize the coordinates of the localization points.Finally,the localization algorithm was verified on 50 rice plants,and 88 % of the localization points met the requirements of the positioning protection area.