The Discovery Of A Specific Object Segmentation Method

Image segmentation has been studied in past fifty years, digital image processing technology has been widely applied in field of computer vision. Image segmentation is the basic and low layer work of computer vision and is also the basic of image understanding and analysis, which takes an important role in computer vision. In order to improve object discernment of the artificial intelligence (AI), object is excellent segmented by image processing technology, and it is object segmentation technology.Specific object segmentation is a problem of image segmentation, and it combines image segmentation with object detection. However, compared with traditional image segmentation, specific object segmentation also has a detection procedure before segmentation, which also uses methods of object detection for detection and makes it more difficulty than traditional one. The early works about object segmentation required some labeling images for training, however it was not helpful for computer discernment new objects. Some excellent classification algorithms were proposed, which based on latent topic model and required large set of training images, it was not helpful for computer discernment efficient. So the study of accurate and efficient specific object segmentation will attract more research's attention in the future.To segment specific object from images, this thesis combines local descriptor and region descriptor to propose accurate image local information descriptor. Then, descriptor and shape model are extracted to generate general model of specific object, which is used for extracting specific object from unknown images. The main results of this thesis are to detect and segment specific object from images. The main works are as follows:1. Local descriptor is used to detect object location. As there are drawbacks of existed local descriptors, local descriptor fusing associated local features is used. Through several matching experiments, the new descriptor is demonstrated as an efficiency descriptor, which gets a higher accurate and guarantees the performance of segmentation.2. Based on over-segmentation, superpixels are first obtained as local regions. Then, descriptor of the local region is designed based on relative color information and contour information. The final local region descriptor is finally generated by considering the new region descriptor and boundary probability model, which is demonstrated to be efficiency descriptor through experiments.3. Combing local descriptor and local region descriptor, a novel model is generated to locate specific object. After that, the object is segmented through fusing local regions.