Research On Discriminative Feature Perception Methods For Scene Segmentation

Scene segmentation,a fundamental and challenging problem in computer vision,aims to assign a category label to each pixel on a scene image.Therefore,scene segmentation requires both precise pixel-wise location ability and accurate pixel-wise identification capability.The rich scene parsing capability enables the applications on scene understanding,autonomous driving,human-computer interaction,augmented reality and other fields.In recent years,benefit from the powerful feature representation capability of deep convolutional networks,fully convolutional network(FCN)algorithms have gradually become the dominant algorithms for scene segmentation.However,existing algorithms ignore the variances of different feature and confuse different feature information,making the feature discriminative degradation and leading to prediction errors.This paper focuses on discriminative feature perception in scene segmentation,and exploit the feature decoupling and feature selection to enhance feature discriminative property,which helps to model scene relationships and segment each element accurately and efficiently.The goal is to extend and improve the FCN algorithm by solving the existing problems including feature differentiation,feature selection and speed-accuracy balance.The main contributions of the thesis are as follows:First,in terms of feature decoupling,for the issue of intra-class inconsistency and inter-class indistinction in the scene segmentation,we propose a boundary and region feature decoupling scene segmentation method.Existing FCN-based algorithms model the scene segmentation as a dense classification problem,ignoring the overall relationship of each class,leading to intra-class inconsistency and inter-class indistinction issue.In this paper,we propose a discriminative feature segmentation algorithm based on the idea of decoupling boundary and region features.The algorithm utilizes multiscale contextual relationships and the channel attention mechanism to extract region features,which enhance intra-class consistency.Meanwhile,the algorithm introduces explicit semantic boundary supervision to extract boundary features,which enlarge the inter-class distinction.The interaction of both features greatly enhances the feature discrimination and improves the segmentation performance.Extensive quantitative and qualitative experimental analyses show that the algorithm is able to solve the intra-class inconsistency and inter-class indistinction issues well,and the feature discriminability and segmentation performance outperform the state-of-the-art methods.Second,in terms of feature decoupling,for the issue of the balance between accuracy and speed in real-time scene segmentation,we propose a detail and semantic feature decoupling scene segmentation method.Current real-time scene segmentation methods tend to sacrifice spatial detail information to improve speed,making it less discriminative on high-resolution features,which in turn leads to a serious decrease in segmentation accuracy.To address this issue,this paper observes that both spatial detail information and semantic information are important to segmentation performance,and both types of information have different requirements for encoding approaches.Therefore,this paper proposes a bilateral segmentation algorithm and an improved bilateral segmentation algorithm based on the idea of decoupling detail and semantic features.The bilateral segmentation algorithm utilizes a shallow branch with more channels to extract detail features and a deeper branch with fewer channels for semantic features.Besides,the algorithm fuses both features efficiently,thus achieving a promising balance between speed and accuracy and improving the segmentation efficiency effectively.In addition,the improved bilateral segmentation algorithm simplifies and refines the original bilateral segmentation algorithm and proposes a bilateral segmentation architecture specifically for real-time scene segmentation.Extensive experimental results show that both algorithm achieves the state-ofthe-art balance of speed and accuracy,and the efficiency outperforms the state-of-the-art methods.Next,in terms of feature selection,for the issue of modelling long-range contextual dependencies for scene segmentation,we propose an implicit guidance feature selection scene segmentation method.Scene segmentation relies on the effective modelling of the contextual relationships.Current scene segmentation methods usually introduce the self-attention mechanism to model long-range relationships.However,the self-attention mechanism has lots of redundant computation,leading to high complexity and limited application.Meanwhile,lots of redundant information pollutes the feature discriminative capability.To solve this issue,this paper proposes a representative graph network segmentation algorithm based on the idea of the feature selection implicitly guided by the similarity relationships.The algorithm learns feature similarity relations and guides the feature selection process,thus selecting the representative features.It can model long-range relationships effectively,reduce the redundant computation of the self-attention mechanism,improve the discriminative representation of features,and substantially improves segmentation performance and efficiency.Extensive experiments show that the algorithm is more efficient than the self-attention mechanism approach and outperforms the state-of-the-art segmentation methods.Finally,in terms of feature selection,for the issue of modelling intra-class and inter-class contextual relationships for scene segmentation,we propose explicit guidance feature selection scene segmentation method.Current scene segmentation methods model contextual relationships without distinguishing their types,leading to relationship confusion and reduced feature discriminability,which in turn results in contradictory predictions.To address this issue,this paper proposes a context prior segmentation algorithm and a conditional classifier segmentation algorithm based on the idea of feature selection explicitly guided by the supervision.The context prior algorithm introduces the constraint of explicit affinity matrix supervision to clarify the intra-class and inter-class contextual relationships,enhancing feature discriminability.By combining the two contextual relationships,the algorithm improves the segmentation performance substantially.In addition,the conditional classifier segmentation algorithm utilizes auxiliary semantic supervision to guide the selection of intra-class features,dynamically generates a sample-specific classifier kernel.The classifier kernel can handle the intra-class variance problem,enhance feature discrimination,and thus improve the segmentation performance.Extensive quantitative and qualitative analysis demonstrates the context prior and conditional classifier algorithms are robust and general.Both algorithms outperform the state-of-the-art methods on multiple scene segmentation benchmarks.This paper mainly focuses on the requirements of practical applications for scene segmentation,exploits the topic of discriminative feature perception in scene segmentation,which is of great importance for promoting the application in practical fields including scene understanding,autonomous driving,human-computer interaction and augmented reality.