Design And Implementation Of Crop Semantic Segmentation System Based On Graph Neural Network

In recent years,China has been vigorously developing precision agriculture.At the same time,agriculture is also the pillar industry of the country.And it is still common in our country to carry out agricultural labor in the traditional way,and there is a huge gap between the level of agricultural technology above and the developed countries.The current computer vision by pixel-level semantic segmentation of crops and other plant roots,stems and leaves,etc.,so that the correct crop information can be obtained from the results,thus playing an effective role in crop generation assessment,crop asset statistics,agricultural information management,etc.And with the rapid development of depth sensor-related technologies and tools,the semantic segmentation models and applications of images incorporating depth information in order to improve the accuracy of segmentation results have meanwhile received extensive attention from the academic community.First,an end-to-end RGBD semantic segmentation model is proposed.For extracting the depth information in RGBD images,a pixel point neighborhood map is constructed by the semantic values of the depth images,and then based on the GCN graph convolutiona l neural network,feature propagation is performed on the obtained initial features to achieve the collection of the spatial dependence of the semantics in the images.Meanwhile,in order to further improve the segmentation results,the model combines the attention mechanism to associate the acquired initial feature channel data as the relevant attention,and achieves the collection of channel dependencies of the acquired features.Finally,the two final features mentioned above are feature fused to obtain the final semantic segmentation results.Secondly,a crop semantic segmentation system is designed and developed.The system integrates the proposed graph convolution-based RGBD semantic segmentation algorithm and the traditional RGB segmentation algorithm with business functions such as real-time semantic segmentation,segmentation file management,data acquisition and data preprocessing.Finally,the proposed algorithm is experimentally analyzed.The GCN graph convolution module and attention mechanism are experimentally tested.Then the experiments were compared with the improved Res Net101 model with 4-channel input,and from the experimental results,the proposed algorithm m IOU and acc results were improved compared with Res Net101.Meanwhile,the semantic segmentation results of the proposed model in this paper were compared with the improved deep learning networks such as mainstream FCN and PSPNet with 4-channel as model input,and all the effectiveness of the segmentation algorithm is verified.The designed crop semantic segmentation system is also verified and tested for each module and function of the system,and the system also has some application value.