Research On Graph-based Visual Scene Representation And Its Applications

As a basic research in the field of computer vision,visual scene representation aims at extracting the hidden structural pattern,describing the internal relationship between visual data and producing the concise and abstract representation of the scene with the use of visual psychology,scene layout,context content and image processing methods.The work has been widely used in the fields of industrial production,military security and life practice,such as intelligent transportation,automatic navigation,environmental monitoring,medical diagnosis,remote sensing analysis,intelligence shopping and so on.Research on scene representation shows that the structured description of visual data plays an important role in the representation of the scene.However,due to the diversity,complexity and variability of natural scenes,it remains challenging for graph-based models to reveal the relationship between visual data in terms of accuracy,efficiency and robustness.Therefore,this dissertation focuses on the study of the relationship representation among scene data from the perspective of path optimization and distance measurement on the graph,and designs algorithms for two hot applications of structure-preserving image filtering and visual attention modeling.The main content of this thesis is described as follows:(1)Research on graph-based path optimization and relation representation between visual data.To solve the problem of description inconsistency and noise sensitivity in the same visual region for graph-based path generation methods,this thesis proposes the smoothest path and smooth-short path optimization methods based on the Gestalt-grouping principles.The method first constructs the smoothest path on the graph to analyze the scene by introducing the Gestalt-grouping laws in human visual perception.Given that the length of the smoothest path between two connected image elements with close spatial distance and similar color feature may be excessively long,the thesis further proposes the smooth and short path on the graph for scene perception.Experimental results show that the smoothest path and smooth-short path algorithms have a strong reliability of detecting the structure information and describing the scene relationship in complex scenes.(2)Research on graph-based path distance measurement and visual data relationship description.Given the fact that the distance measurement methods applied in the field of scene representation are easily affected with the bias problem caused by limitations such as non-linear data structure,varying illumination and blurry boundaries when representing the relationship between visual data,this dissertation proposes a path distance metric based on the bottleneck analysis.By taking the semantic information and topology structure between each pair of nodes into consideration,the thesis introduces the path bottleneck detection and analysis method based on the random walk model to form the relationship representation between visual elements.The experimental results demonstrate that the bottleneck detection method can not only maximize the differences between-classes while minimizing them within-class,but also preserve the key dissimilarities between visual elements in the same region.(3)Research on formulating structure-preserving image filter in natural scenes.To deal with the problem of edge blurs caused by the similarity between the scene structure and high-contrast detail in terms of gradient,the thesis proposes an automatic scale-awareness based structure-preserving image filtering method.The method first introduces the clustering distance transform method to aggregate high semantic-consistency visual elements,and then establishes the cooperative filtering model to guide the smoothing process.The model is formulated based on the study of the decision information from neighbors and trust mechanism to integrate the clustering distance transform and bilateral filter together.The experimental results indicate that the collaborative structure-preserving image filtering model is robust to structure keeping and high-contrast detail smoothing.(4)Research on modeling the visual attention mechanism in natural scenes.To tackle the problem of inconsistently highlighted salient objects in the image due to factors of arbitrary scales and uncertain distributions of scene regions,uneven illumination and cluttered background,this thesis proposes a structural representation based salient region detection algorithm.The method first establishes the path by introducing the Gestalt-grouping theory,then applies the Laplacian analysis on the path to measure the connectedness,and finally,combines the background connectivity priori with the appearance contrast cue to define the saliency.The experimental results prove that the proposed saliency detection model is capable of uniformly highlighting salient regions and suppressing non-salient ones.