Research And Application Of Image Semantic Segmentation Based On Deep Fully Convolutional Networks

Semantic segmentation refers to assigning a category label to each pixel in an image so that it can not only segment the region,but also mark the content.As more and more visual applications are in urgent need of accurate and efficient segmentation technology,such as autonomous driving,medical image analysis,video surveillance,augmented reality,etc.Im-age semantic segmentation,which is known for precision and fineness,has received more and more attention in the field of computer vision and machine learning.In general,semantic segmentation combines classification and location to solve the problem of pixel-level prediction.However,how to balance the high-level abstract classifi-cation and the low-level precise location is the difficulty faced by semantic segmentation.This paper aims to improve the hierarchical feature fusion through attention mechanisms to obtain better image semantic features,and construct the object boundary detection model to refine the segmented object boundaries.The main work and innovations are reflected in:(1)Focused on the issue that the semantic segmentation based on fully convolutional networks is prone to local area perceptual errors when dealing with objects with complex appearance,a method based on feature fusion guided by attention mechanisms is proposed.The method uses the high-level semantic information with the strongest semantic con-sistency constraint to fuse the hierarchical features of different scale contexts from top to down.The attention module provides fusion guidance and constrains the semantic con-sistency of the fusion feature to obtain the best prediction.The experimental results on the PASCAL VOC 2012 and Cityscapes datasets show that the improved model can capture rich context information and obtain the image feature with more intra-class semantic consistent,which has obvious advantages over similar methods.(2)Focused on the issue that semantic segmentation methods based on full convolutional networks often lead to the problem of blurred object boundary in segmentation results,a method based on object boundary detection is proposed.The method trains the detection model with a new data set containing object boundary labels transformed from the existing data set to obtain spatial information about the object boundary.Through the mask module guidance,the semantic features extracted by the segmentation model are combined with the spatial features extracted by the detection model,and the boundary information of the object is used to refine the result of semantic segmentation.The experimental results on the PASCAL VOC 2012 and Cityscapes datasets show that the model based on boundary detection can obtain more clear and detailed object segmentation results.(3)Focused on the specific application of clothing parsing,a clothing parsing method based on a garment encoder predicting and constrainting combination semantics is proposed based on the above research,with a corresponding dress collocation system.The method introduces a garment encoder branch derived from the end of the full convolution network which was used to predict the combinatorial preference of the garment items,thereby filter out the indeterminate labels.A fully-connected conditional random field is used to improve the segmentation quality as the post-processing step to correct the prediction.The experi-mental results on the Fashionista and CFPD datasets confirm the feasibility of the method.