Contour Map Reconnection Research Based On MSDAE

The image is one of the main means for human accessing to informations. Especially in today's era of big data, image plays an important role both in people's daily life and scientific research. However, image may lose part of information by many factors when we record or disseminate it. Especially in the contour map, it always occur fractures for it specific graphical structure when we spread it. So the study on reconnect contour map is of great significance. Most of traditional reconnect algorithms are of bad commonability and effect, so how to improve the versatility and the effect of the algorithms has become a key issue. By advancing in technology and rapid development of some learning algorithm, it becomes possible to handle the problem in an intelligent way.This thesis studies on how to deal with the disconnection problem in contour map. We focus on designing a new model from auto encoder which is suitble for contour map reconnection problem. The main works is as fellows.1. There are three classical deep learning models which are called conventional neutral network, deep believe network and auto encoder. By researching and analysing the three models on aspects of model characteristics and application scenarios, we find out auto encoder is the most suitable model for the problem we face to. After find the model we need to expand the idea of application by seeing it as a recovery model instead of a feature training model.2. Based on the improvement above, a new contour map reconnection model which is called multi-direction stacked denoising auto encoder(MSDAE) is given. There are two main part in this model, training part and testing part. The training part contains four groups of multilayered structures, which are the new strcutures improved from auto encoder. This four structures use samples in different directions. It makes each model has a strong reconnect ability in dealing with samples in the direction. This method facilitates the synthesis of final result. The testing part is a composite structure; a composite method is given for the final result. First we combine the results of the four parallel models and the chose the best one as the optical result. Then the final result will combine the optical result and the others mentioned above.Different maps are used to test the model MSDAE. We also campre the model with the minimum points method by using same test set. Experiment shows that the model has a good accuracy in dealing with the contour map reconnection problem.