Research On The Optimization Of Boundary In Semantic Segmentation Based On Deep Learning

Semantic segmentation aims to give pixel-level classification result to the content of an image,then the image will be divided into several specific areas with common semantic information.These extracted regions of interest will be used for subsequent image process or content understanding.So,semantic segmentation is widely used in related vision tasks,such as autonomous vehicle,medical image analysis and scene understanding.In recent years,Convolutional Neural Network(CNN)-based approaches have achieved state-of-the-art performance in many vision tasks.Supervised deep learning algorithms have achieved incredible performance better than most traditional algorithms.However,in a large application scenario,the precision of edge segmentation still can't meet the requirement.In the field of natural image segmentation,automatic driving,various kinds of live video applications and function of beauty all have a high requirement for segmentation accuracy on edge.With the release of a large number of high-precision annotated datasets,the improvement of the boundary segmentation precision of natural image segmentation has becomes a research hotspot in the field of computer vision recently.In the field of medical image process,accuracy boundary segmentation of anatomical structure is required to provide strong support for subsequent clinical diagnosis.Therefore,edge segmentation has always been the focus of its research.Considering above problem,in this paper,we carried out a detailed study on the boundary segmentation optimization algorithm in the field of image semantic segmentation.On the task of natural image segmentation,we proposed a feature decoupling module guided by edge information and a multi-level feature-based boundary-pixel pointwise refinement algorithm.The current network feature extraction process lacks explicit modeling on boundary and interior pixels,in order to further improve the consistency of the internal features of the object extracted by the network and the detail of boundary features,we proposed a feature decoupling module guided by edge information,which decouples feature into boundary feature and inner feature.The module improves the ability of feature extraction of the network.Simultaneously,the existing segmentation network encoder inevitably uses downsampling operations and coarse bilinear interpolation or deconvolution to upsample,which leads to the loss of high frequency information.So we propose the pointwise refinement algorithm,it represents pixels on boundary with multi-level features extracted by encoder,then uses a multi-layer perceptron refinement structure to refine the coarse prediction of segmentation network,so the segmentation of boundary pixels is improved with a very small amount of calculation.The methods above effectively improved the performance of multiply state-of-the-art segmentation backbones on the high-precision annotated natural image dataset Cityscapes.Finally,in the field of medical image segmentation,we propose a data augmentation framework based on Transfer Learning,which solves the problem of poor edge segmentation performance due to lack of high-quality annotated data,we also verified our work on the medical image segmentation tasks including heart segmentation and human abdominal segmentation.