| Currently,the complete development process of coconut is not fully understood for the time being.For traditional planting methods,it is often necessary to observe the structure inside by destructive means at multiple stages.For this reason,it is difficult to meet the demand for variety diversification and high germplasm standards due to the low efficiency of the previously mentioned method,which seriously hinders the formation of differentiated competitiveness of coconut products.Under such circumstances,it would be a very important and challenging task to combine non-destructive observation and artificial intelligence technology to describe the development of internal organs qualitatively and quantitatively.To deal with the above-mentioned problem,a non-destructive coconut CT image is built and an improved Deeplab V3+ network segmentation model and a 3D reconstruction method for optimizing mining methods are proposed in the paper,to reveal the growth and development of coconuts by using 2D semantic information graph and 3D simulation forms,as well as corresponding quantitative data and to assist researchers and planters in making scientific decisions.The main work is as follows:(1)Build the dataset of coconut CT image.Utilizing the non-destructive characteristics of CT scans and targeting multi-variety and multi-stage coconuts,the Coconut Research Institute and the Radiology Department of Municipal Hospital jointly worked to establish a CT image dataset of coconuts containing different performance traits and conduct professional annotation,providing valuable image resources for the image field and coconut research.(2)An improved semantic segmentation model based on Deeplab V3+ is proposed.To handle the problems of low target resolution,unclear linear features,and mutual interference between organ regions in coconut CT images,a new model based on Deeplab V3+ infrastructure has been designed for improvement,which replaces the original ASPP structure with a densely holed spatial pyramid pooling module and introduces a CBAM attention mechanism to solve the problem of information loss caused by sparse sampling while better capturing global features.Subsequently,a residual refinement module is embedded in the decoder to optimize the boundary information between organs,and multiple model comparison verification and ablation experiments are conducted eventually.The findings indicate that the improved segmentation algorithm has higher accuracy in the face of the polymorphic coconut organ CT image and a single type of organ can be extracted and corresponding quantitative data can be obtained through the semantic segmentation graph of coconut.(3)A 3D reconstruction method of coconut fruit based on optimized sampling method is put forward.To solve the problems of time consuming and poor contour details in the reconstruction process of Alpha-Shape based on coconut point cloud data,KD tree is introduced to establish an associated topological relationship between massive unordered discrete points in a coconut point cloud set,thus the value of α can be determined based on the density of point clouds within the region.What’s more,the ability to quickly find corresponding correlation points during contour rendering greatly accelerates the reconstruction speed,while making the detailed information of the surface more complete,making the reconstruction target closer to the actual object.Comparative experimental results demonstrate that the optimized method performs better in the 3D reconstruction of coconut,and the reconstructed 3D model can also obtain the desired target structure data in this dimension.(4)The intelligent segmentation and reconstruction system for coconut is designed and realized.By encapsulating the proposed improved algorithm model in the background,the user calls the back-end model through the intuitive front visual interface to obtain the required data,which provides an easy-to-operate and practical auxiliary tool for the analysis and evaluation of coconut fruit growth. |
PDF Full Text Request