Research On 2D Object Shape Matching And Generation

Shape analysis is one important branch in computer vision and image processing,it has been widely used in engineering areas,such as object recognitions,medical diagnosis and target retrieval,etc.In this thesis,we research on recent progress of shape analysis,and focus on two sub-problems:shape matching and generation.The main work of this thesis contains the following two parts:(1)Develop a new shape descriptor which employees the angle sensitivity,and use it for shape matching problems.(2)Use Variational Auto-Encoder for shape generation,and optimize the initial parameters of its networks.Shape matching is the process of abstracting object shape geometric features for recog-nizing and retrieving same shapes from the database.Most shape matching algorithms can be divided into two categories:contour-based and region-based methods.In practice,there are many factors that can influence the matching and retrieval accuracy,such as scaling,rotation and noisy,etc.At present,most shape matching methods are focus on the correct retrieval accu-racy,and do less analysis for matching time.However,with the rapid development of modern imaging devices,the numbers of images are growing up quickly,the speed of shape matching and retrieval becomes more important.Getting a good balance between accuracy and speed is a challenging problems.In this work,we propose a new shape descriptor DAPD(Dynamic An-gular Partition Descriptor),which contains the angle sensitivity from observation view ranges.DAPD is based on sampling of shape contour points,and it has low computational complexity so it can be used for fast shape matching.In experiments,it shows that it can speed up to 13X compared with some other classical shape matching methods.Meanwhile,we optimize the al-gorithms of measurements,and use experiments to show that it can be robustness to scaling,rotation and noisy.Shape generation is a new topic in imaging processing areas.By abstracting geometric features from existing shape datasets(such as shape contour points or shape gray-scale images),it forms a latent shape space,and we can use these features to generate new similar shapes,so it can assist people to draw new images and so on.Most shape generation methods can be divided into geometry-based methods and statistic-based methods.It's very hard to generate high quality shapes,due to the high dimension of shape images and uncertain errors by handwritten draw-ings.In this work,we apply VAE(variational auto-encoder)method for research on shape latent space generation.It uses encoder to encode shape images into latent variables and use decoders to generate new shape images.By setting object functions and training the neural networks of encoder and decoder,it can make the distribution of generative model similar to true shape dis-tribution.It establishes a shape probability space based on the maximum likelihood,and make features presented by latent variables.Because it is based on statistics,so this method can well reduce the impact of uncertain accident factors.Through encoders and decoders,VAE makes a bridge between low-dimension latent variables and high-dimension shape images.Fi-nally,we use multiple datasets to do various experiments,and analyze the experimental results of shape generation.Finally,we apply the shape matching and generation algorithms for a tailored case,and make new high-dimension clothing shapes by low-dimension latent variables.