| Rapeseed is the largest oil seed crop in China.It has high application value and potential.The accurate three-dimensional reconstruction of rapeseed is beneficial to the breeding research and field production management of rapeseed.In order to solve the problems of high price and low precision in 3D reconstruction of rapeseed,this paper proposes a method of phenotypic reconstruction of rapeseed under limited perspective,designs an algorithm for depth estimation and new view generation of rapeseed,and applies it to 3D reconstruction field.It solves the problem that 3D reconstruction accuracy is not high in finite view Angle image.The algorithm designed in this paper was first trained and tested on public data sets to verify its effectiveness,and then applied to rape data,and further extended to the three-dimensional reconstruction of rape plants.The main contents and conclusions of the study are as follows:(1)The methods of plant phenotypic reconstruction were summarized and analyzed.By comparing the advantages and disadvantages of the method,it is helpful to reduce the cost to determine RGB image as the research object.Firstly,two algorithms are designed to process RGB images.The two algorithms are unsupervised depth estimation algorithm and new perspective generation algorithm based on neural radiation field.The results of the improved algorithm on the public data set are better than the existing algorithm.Then,the results were applied to the 3D reconstruction of rape plants,and the high quality 3D reconstruction was realized under the condition of few view angles.(2)An unsupervised depth estimation algorithm based on monocular video is designed.Its network structure is composed of depth estimation network and pose estimation network.In the depth estimation network,a feature extraction network TC-Net based on Transformer and CNN is designed to extract more abundant image features when some image feature points are less.A feature enhancement module is proposed to enrich the capability of feature extraction.Photometric reconstruction error is used as the main loss function in the network.In addition,in order to further improve the accuracy of depth prediction,loss functions such as Multi-Scale feature loss function and PSNR loss are added to the loss function.The results are as follows:=4.872,7)2)=0.179,(70)7)=0.116,0)7)=0.918 and(6(8(8(6(8=0.907 on KITTI data set.The results on the SCALED dataset with fewer features were=14.673,7)2)=0.194,(70)7)=0.146,0)7)=2.895 and(6(8(8(6(8=0.957.Further ablation experiments were conducted to verify the influence of each module on the algorithm.Experimental results show that the algorithm proposed in this chapter is competitive with the current mainstream unsupervised monocular depth estimation algorithms,which lays a foundation for the design of new view generation algorithm in Chapter 4.(3)A new view generation algorithm based on depth map and nerve radiation field is designed.Based on the traditional NeRF network structure,this algorithm integrates deep information into the NeRF network in response to the poor prediction accuracy of the NeRF network model in the case of fewer views and poor rendering of new views.Its significance lies in the fact that it can generate high-precision new perspective image with less training perspective.In terms of accuracy,this algorithm achieves good results.On Scan Net dataset,this algorithm was compared with NeRF method,and a new view was generated,which was clearer.Taking,andas quantitative evaluation indexes,the accuracy of this algorithm was as follows:=31.68,=0.956 and=0.194,all of which were better than NeRF algorithm.In order to verify the effectiveness of the algorithm,experiments are carried out on multiple data sets,and the results are good.It laid a foundation for realizing three-dimensional phenotypic reconstruction of rape plants from a limited perspective.(4)The two algorithms were applied to the 3D reconstruction of rapeseed plants,and the 3D reconstruction under the condition of few view angles was realized.The video data of rape was collected and preprocessed to form a sequence image of rape,and the data set was made.The unsupervised depth estimation network is applied to the data set,the depth estimation is carried out,and the experimental results show that the unsupervised depth estimation algorithm proposed in this paper is better than the traditional algorithm in predicting the depth map.The new perspective image generation algorithm based on neural radiation field was applied to the rape plant data set,and the depth map predicted by the unsupervised depth estimation algorithm was added to the training of this algorithm.The addition of depth information made the algorithm train faster and generated high quality new perspective images.For rape plant3D reconstruction,given a certain Angle of view of rape plant image,combined with the generated new Angle of image 3D reconstruction,experimental results show that compared with direct reconstruction,the method proposed in this paper has better 3D reconstruction effect,can achieve 3D reconstruction under fewer angles,and its application is more extensive.On the basis of 3D reconstruction,the phenotypic parameters of rapeseed were measured and analyzed.Leaf width,plant height and rhizome diameter were measured by 3D and manual methods,and the determination coefficient~2 and root mean square errorwere used as evaluation criteria.The experimental results are as follows:correlation coefficient~2=0.993,root mean square error=0.53(88);Correlation coefficient~2=0.996,root mean square error=0.95(88);Correlation coefficient~2=0.931,root mean square error=0.25(88).The experimental results showed that the phenotypic parameters measured by 3D reconstruction were close to those measured by hand,and a good result was achieved.In conclusion,we designed a depth estimation algorithm and a new view generation algorithm,and provided a set of low-cost and efficient methods for 3D reconstruction and phenotypic monitoring of rape and other crops in a limited perspective. |
PDF Full Text Request