Super-pixel Constraint Based Road Semantic Segmentation

Semantic segmentation is a fundamental question in the field of computer vision.Semantic segmentation is the preposition task of many researchers,including pathological analysis,autonomous driving,character reconstruction,and so on.Many scholars are committed to improving the accuracy of semantic segmentation.Since deep learning is applied to semantic segmentation,the accuracy of semantic segmentation has been dramatically improved compared with traditional methods.Many deep learning-based semantic segmentation frameworks have been proposed to solve the task.Given an input image,the goal of semantic segmentation classifies every pixel of the image,and the output is the same size as the input image.Due to the nature of the subject itself,the object boundary is challenging to be segmented accurately.On the other hand,due to the variety of the scene and the difficulty of acquiring data.Alarge number of weakly supervised and semi-supervised deep learning methods have emerged.However,the specific road image segmentation task faces class unbalances,so the classification accuracy is too low for categories with small pixel proportions.In order to improve the learning ability of the semantic segmentation model and improve the accuracy of semantic segmentation,the existing methods try to solve the problem through designing many complex deep frameworks.In addition,the existing methods lack the correlation research and application of the information between pixels on the image.To solve the above problems,this paper proposes a semantic segmentation method based on superpixel constraints.The method is mainly a plug-in semantic segmentation module with positive feedback.Specific work is as follows:·This paper proposes a multi-task semantic segmentation method with a superpixel segmentation branch as an auxiliary branch.The superpixel segmentation is a correlation task with the semantic segmentation so that it can effective improve the semantic segmentation precision.·Based on the above multi-branch framework,the paper proposes the correlation constraint between pixel features based on the superpixel segmentation.Specifically,the paper considers the similarity of pixel features in the same superpixel,the difference of pixel features in the different superpixels,and the correlation of pixel features of the same segmentation label among multiple superpixels to constrain the network.·The paper proposes a general-purpose module for semantic segmentation methods.The general module is appropriate for the encoder-decoder structures.It can significantly improve the pixel segmentation accuracy of small categories caused by sample imbalance through carefully designed category weights.This paper mainly conducts experiments on the scene segmentation task,that is to say,the road image segmentation task.The paper experiments on the CityScapes dataset.The ablation experiments and the compared experiments with advanced semantic segmentation are designed to evaluate the proposed method through qualitative and quantitative experiments.The experiments verified that the method in this paper has a certain improvement on semantic segmentation,especially for the problem introduced by category imbalance,and achieves stateof-the-art performance.