| The vanishing point(VP)in the image provides important information in the 3-D scene.It has been widely used in the fields of computer vision such as 3-D reconstruction,robot navigation,and so on.At the same time,it also provides great help for humans and artificially intelligent algorithms to analyze image perspective relationships and learn photography skills.Therefore,the research about the detection of vanishing points is very meaningful.Traditional machine learning based methods usually take account of domain knowledge,decompose the VP detection task into multiple small steps,and optimize them individually.However,such methods typically rely on hand-crafted features,and the errors of each step can be accumulated,which decreases the accuracy of VP.Whereas convolution neural network(CNN)is an effective model for detecting vanishing points in images,which can automatically learn the feature representations of vanishing points from images and avoid the disadvantages of traditional methods.However,the success of CNN heavily relies on a massive amount of training data to ensure high accuracy.Without sufficient and balanced training data,the obtained CNN-based VP detection models can be easily overfitted with less generalization.In view of the existing problems of the existing VP detection algorithms,this thesis has launched research work,including:1.Aiming at the problem that traditional methods with individual optimization cause error accumulation,in this thesis,the three functions of edge detection,line fitting and VP prediction are implemented as an endto-end neural network,through which the mapping from image space to edge space,then to line space and finally to VP space is learned.2.Aiming at the problem that current CNN methods rely on a large amount of data and are difficult to learn the geometric relationship in the image,this thesis acknowledges that a VP in the image is the intersection of 2-D projections of multiple parallel lines in the 3-D scene and treating this knowledge as a geometric prior,and proposes a prior-guided residual line-shaped convolutional network for VP detection to reduce the dependence of CNN on training data by introducing prior knowledge.3.In order to show the benefits of VP detection for aesthetic tasks,this thesis also implements an image retrieval algorithm based on vanishing points,which calculates the similarity of VP geometric information and retrieves the image closest to its perspective relationship from the collected images with excellent aesthetic quality.Finally,an interactive system is implemented for users to perceive perspective aesthetics through vanishing points.Our extensive experiments on public datasets show that the proposed geometric prior-embedded VP detection network improves generalization performance and inference speed. |
