Research On Semantic Image Segmentation Algorithm Based On Deep Learning

Semantic image segmentation,which aims to assign a label to each pixel in an image,is a fundamental task in computer vision,and it has a wide range of applications.In recent years,semantic image segmentation algorithms based on deep learning have received extensive attention because of their faster speed and higher accuracy.However,due to the large number of down-sampling layers in deep model,the segmentation result usually performs poorly in the region near object boundary,of which there is currently no general quantitative evaluation metric to measure the quality.In order to solve these two problems,this thesis introduces region-based evaluation metric to solve the latter problem,and proposes a modified loss function to improve the performance of segmentation model.The main contents are as follows:Firstly,current methods for semantic image segmentation are summarized and the segmentation model DeepLab used in this thesis and common evaluation metrics are introduced in detail.Secondly,in view of the fact that conventional evaluation metric cannot quantitatively measure the quality of segmentation result in the region near object boundary,this thesis proposes region-based one.By using the ground truth of semantic image segmentation task,this thesis presents a method to extract edges efficiently and accurately.By defining the distance from pixel to the edge of object,an algorithm for quickly calculating the edge region is proposed.Based on this,three types of region-based evaluation metric are introduced.The experimental results show that the region-based evaluation metrics proposed in this thesis,can quantitatively and accurately measure the quality of segmentation result in the region near object boundary,and based on the analysis of these metrics,this thesis points out the problem existing in the DeepLab model.Finally,this thesis proposes a modified loss function to alleviate the problem that segmentation result performs poorly in edge region.Based on the analysis of the previous experimental results,this thesis gives the definition of hard example in semantic image segmentation task,and uses multi-task learning mechanism to propose a loss function that focuses on hard example learning.The experimental results show that the proposed loss function can bring about 1% improvement compared to the current mainstream cross entropy loss function,in terms of region-based evaluation metric,4% improvement.Besides,other models use the modified loss function also outperform their counterparts in terms of both conventional and region-based metric.