Research On The Semantic Segmentation Of Complex Scene Image Of Field Based On Fully Convolutional Networks

With the continuous development of deep learning,depth convolution neural network is increasingly applied in the area of computer vision,so that it has a broad application prospects in the field of intelligent agricultural machinery operations.In order to get the environmental information of field more effectively,images taken in Northwest A&F university experimental fields were selected as research objects.Deep learning technology,combining with supervised learning method,so as to realize the semantic segmentation of complex scene image of field.The main contents and conclusions are as follows:(1)Design of Fully Convolutional Networks model.According to lots of kinds in the field that is susceptible to light,weather,temperature and many more,the traditional segmentation algorithms can not solve the above problems effectively that is why the Fully Convolution Networks(FCN)was designed for the semantic segmentation of complex scene image of field.The semantic segmentation of pixel level of field image was realized by deconvolution.(2)Semantic segmentation of complex scene image of field based on FCN.According to the overfitting problem of FCN,the data enhancement algorithm was studied.Then in order to solve the problem of the depth learning network training for a long time and convergence,this paper adopted two stage for training model.Finally,FCN-32 s,FCN-16 s,and FCN-8s are adopted to establish the structure of upsample,and the semantic segmentation results are analyzed.The experimental results show that the results of FCN-8s is the highest in three network structures,in which the pixel accuracy can reach 90.87%,and mean IU can reach 75.52%.According to the fact that the field with surfficient illuminition and insurfficient illuminition situation,the model is tested with the surfficient illuminition and insurfficient illuminition data sets.The experimental results show that the model has good adaptability and stability to the field with surfficient illuminition and insurfficient illuminition situation.(3)Optimization of semantic segmentation model for complex scene image of field.As the traditional activation function is not sufficient for the data and the robustness to the noise is poor,we studied the model of the rectified linear unit(ReLU)and the exponential linear unit(ELU)activation function in this paper.The experimental results show that the ELU activation function can accelerate the training speed of model and has better noise robustness.We adopted Batch Normalization(BN)to make the FCN training process more stable and play a regularized role.Support vector machine(SVM)and maximum multinomial logistic regression(Softmax)are adopted to establish models which suitable for semantic segmentation model for complex scene image of field,and the semantic segmentation results are analyzed,the experimental results show that the results of SVM is better than Softmax.The optimized model,in which the pixel accuracy can reach 91.86%,and mean IU increases to 76.84%.The experimental results show that optimized model is more suitable for semantic segmentation of complex scene image of field.In conclusion,the ELU activation function,model two-stage training,and SVM classifier are adopted to realize the semantic segmentation model of complex scene images based on FCN,in which the validity of the model is verified by experiments,which can make pixelwise prediction with original image.Besides,it provides a basis for the research of semantic segmentation of complex scene image of field.