Research On Image Semantic Segmentation Method In Farmland Scene

China is a big agricultural country and its agricultural economy is of great importance.Helping agricultural machinery to better understand the surrounding environment plays a crucial role in promoting the development of precision agriculture.And the key research content in scene understanding is the semantic segmentation.This paper takes farmland scene images as the research object,and acquires,enhances and processes the farmland image dataset.According to the fact that the farmland scene is easily influenced by the environment,the semantic segmentation of the image in the farmland scene is realized by combining the conditional random field with fully convolutional networks and adding the edge region constraint.The main research contents and conclusions of the article are:(1)Get the datasets for semantic segmentation of farmland images.In view of the fact that the existing datasets lack complete field samples,by referring to the existing mature common image semantic segmentation datasets,firstly,the photographing location,time,method and object of the farmland scene are designed,and the image acquisition work is completed.Then,according to the information content of the acquired farm scene image,the dataset is enhanced by flipping transformation and translation transformation in the horizontal direction,so that the number of images becomes 8 times the original to prevent over-fitting.After the adjustment of dimensions and ground truth,a complete set of dataset for image semantic segmentation in farmland scenes was established.(2)Semantic segmentation model of farmland based on fully convolutional networks.After the farmland scene image dataset is constructed,under the framework of caffe,firstly,the model training and test operations are completed based on the initial FCN model of the VGG-16 network structure.Then according to the criterion of image semantic segmentation accuracy,the semantic segmentation results generated by three different skip architectures of FCN-32 s,FCN-16 s and FCN-8s are judged and analyzed.Lastly,FCN-8s with good segmentation effect in the farmland scene was determined as the research basis for follow-up work.(3)Combining the CRF-RNN farmland semantic segmentation model and optimization.For the problem of inadequate segmentation and lack of connection between pixels in the FCN model,the conditional random field as recurrent neural network(CRF-RNN)is used to enhance the context information between pixels of the farmland images,and according to the actual scenes of the farmland,the existing model is improved by the edge constraint,and the PA is 90.05%,MIoU increased to 73.45%.Finally,the experimental results were judged and analyzed according to the accuracy of image semantic segmentation,light intensity and crop density,and the effectiveness and strong applicability of the farmland model were determined.In summary,this paper takes the acquired farmland image dataset as the research basis,almost achieves the image semantic segmentation in the farmland scene,and verifies its validity and applicability,and provides a theoretical basis for the agricultural machinery to perceive the realization of the surrounding environment.