Research On Semantic Segmentation Algorithm Based On Fully Convolutional Neural Network

Image semantic segmentation is a key technique for image understanding,which semantically segments an image by assigning a category label to each pixel in the image.Based on the research and analysis of the existing fully convolutional neural network semantic segmentation algorithm,this paper explores the combination of multiple strategies to improve the segmentation accuracy,such as missing detail information,missing context information,and multi-scale object adaptation.The following three aspects of research work have been carried out:1.In this paper,the problem of image detail loss caused by the whole convolution network in the pooling and downsampling process is adopted.The convolution is used to replace the standard convolution in the full convolution network,thus effectively expanding the receptive field and retaining the image details information.Simulation experiments show that the algorithm uses more context information in feature mapping to obtain more detailed features,which effectively improves the accuracy of image semantic segmentation.2.In this paper,an image semantic segmentation method based on enhanced dense atrous spatial pyramid pooling algorithm combining global context is proposed.The atrous spatial pyramid pooling algorithm does not have enough features to deal with high-resolution input images,and the acquired receptive field is not large enough.The algorithm combines the input features of the low expansion rate sampling layer and its output features as the input of the subsequent high expansion rate sampling layer on the basis of parallel sampling;in addition,it also performs a global average pooling on each output feature to integrate into the global context.Information to obtain useful context information in a multi-scale feature map.Simulation experiments show that the algorithm effectively improves the segmentation ability of objects of different scales.3.In this paper,the convolution condition random field algorithm is used to optimize the image semantic segmentation output.Because the traditional full connection condition is inefficient with the airport learning,this paper adds the conditional independence hypothesis to its framework,represents most of the reasoning as a convolution operation,and designs the overall conditional random field as a circular convolution network module,embedded in the network for end-to-end training.Simulation experiments show that the algorithm not only improves the speed of reasoning but also improves the segmentation accuracy.