Research On Object-level Context Constrainted Semantic Edge Detection Algorithm

Semantic edge detection(SED),i.e.,simultaneous localization and classification of target object contours within an image,has important application values in fields such as medical image analysis,robot visual perception,and automatic driving.Existing semantic edge detection algorithms generally adopt Convolutional Neural Network(CNN)as backbones to extract multi-scale features.After theoretical and experimental analysis,we found that existing SED algorithms had difficulty in coping with the conflict between the two subtasks of contour localization and semantic classification in terms of semantic and detailed feature requirements.At the same time,existing algorithms have difficulty in meeting the different demands of contexts for classifying different contour points.On the other hand,the criterion used to evaluate semantic edge is one-sided,and it cannot comprehensively evaluate the advantages and disadvantages of different semantic edge detection algorithms.To solve the above-mentioned problems,this thesis proposes new SED algorithms and evaluation criterions.Specifically,(1)We propose a semantic edge detection algorithm,All-Hi S-In Net(All-HigherStages-In Network),which core structure is All-Hi S-In ACA(All-Higher-Stages-In Adaptive Context Aggregation).All-Hi S-In Net can adaptively aggregate intra-and inter-object semantic features from any high-stages for low-stages edge features,to keep details while aggregating sufficient semantics.The excellent performance of AllHi S-In Net is verified through sufficient comparison with existing SED algorithms based on other multi-stage feature fusion structures.(2)Further,with All-Hi S-In Net as the base network,we propose OFCENet(Object-level Feature Consistency Enhancement Network).It uses a two-branch structure to group details and semantic features separately.It contains a newly proposed Object-level Semantic Integration(OSI)module to enhance the consistency of semantic features within the same objects.In addition,we propose an Inter-layer Complementary Enhancement(ICE)module to perform feature enhancement on each branch.Comparative experiments verify that the OFCENet algorithm has better semantic edge detection performance than existing methods.(3)To evaluate semantic edge detection methods more comprehensively,this thesis extends existing semantic edge quality criterions in aspects of edge thickness,continuity,and smoothness.We validate the effectiveness of the new metrics via user study.Finally,the extended evaluation metrics are applied to compare existing SED algorithms with the proposed OFCENet algorithm,in which the superiority of the OFCENet algorithm is presented in a more comprehensive way.