Research On Shape Detection And Extraction In Complex Scenes Using Yin-yang Discrete-point Computing

Shape detection and extraction is one of the key techniques on the field of object detection, which to some extend represents the developing level of computer vision, and demonstrates its influential role with widespread application prospect. However, detecting and extracting shapes quickly and robustly have been facing huge difficulties and severe challenges coming from complex scenes where objects located. Yin-yang discrete-point computing is a new image analysis method proposed in this dissertation for overcoming such difficulties and challenges. In this dissertation, line shapes and planar shapes in complex scenes, and line shapes with complex structures, are taken as the subject investigated, and how to detect and extract such shapes quickly and robustly is taken as the objective studied. Related theories and techniques researches are developed in several aspects of the Yin-yang discrete-point computing, including research requirement, method proposing, basic application, modified application and enhancing its adaptability.At the beginning of this dissertation, the problems faced by object detection in complex scenes, shape extraction, shape matching, and natural handwriting line segmentation are analyzed, and two common problems, including unstable extracted low-level features and difficulties with organizing them, are concluded. While the methodology of granular computing, including building a problem’s granular structure and "solving the problem in its granular structure", offers a set of theoretical framework for solving these two common problems. So starting with the philosophical thought and methodology of granular computing, this dissertation infers feature discrete-point computing and yin-yang discrete-point computing. Feature discrete-point computing is a detail realization and application of granular computing on image analysis, while yin-yang discrete-point computing is one kind of discrete-point computing handling shape detection and extraction in complex scenes with whether the gray value of local region change saliently or not as a salient feature.Aiming at problem of unstable extracted low-level features, this dissertation proposes the Yin-yang discrete-point sampling model. This model is a type of fixed grid sampling using adaptive local parameters, whose sampling model is built according to Weber’s law, and can be adapted to illumination changing automatically. Restraint on repeated stimulus is reflected in this sampling model, so it is not sensitive to interfere from noises and textures, and with the ability of enhancing objects and inhibiting backgrounds automatically. Boundaries of planar objects and slender curve structure are not filaments or parallel lines depicted by single connected edge pixels, but are belt-shaped regions depicted as clusters of discrete points, which bring great convenience for subsequent yin-yang discrete-point grouping.Aiming at problem of organizing extracted low-level features, this dissertation proposes Yin-yang discrete-point grouping methods. There are two kinds of grouping methods, including shape extraction using centerlines grouping and shape detection using contour template matching. The former one by analyzing discrete-points’features in different belt-shaped regions of yin-yang discrete-point maps, produces a three-level detection system using feature points, line segments and centerlines, which can extract centerlines with their types quickly, smoothly and accurately from these belt-shaped regions, and achieve the final shape results by clustering these centerlines with Gestalt rules. The later one uses yin-yang discrete-point maps as an intermediate representation, to enhance the stability of the scene maps, matches hand-sketches with regions in yin-yang discrete-point maps, to reduce the difficulty of the matching calculation, and according to the different situations of differences between shape wanted and contour template, designs three template optimized strategy including selecting, transforming and estimating. Integrated using those grouping methods, the requirement which want to detect and extract large scale shape with affine transforming can be met.On the aspect of basic application, the use of yin-yang discrete-point computing in the Trouble of Freight Car Detection System (TFDS) is introduced with emphasis. There are some geometric transformations and large nonlinear deformations existing with the photos taken by the TFDS, and the requirement for a image intermediate representation is stability, which keep it away from interferences of illumination changing, spot noise and geometric transformations. Our yin-yang discrete-point sampling model satisfies this requirement very well. This dissertation utilizes sampling and all kinds of grouping methods to research the problems such as similar components classification and wheel locating in the TFDS. The results of noise testing in the laboratory and more than three years field operation prove that the methods above are accurate, real-time and robust.On the aspect of modified application, this dissertation first reforms the yin-yang discrete-point sampling model, and makes even gray area in an image more salient in an enhanced yang discrete-point sampling map, then combined with the real applications’ requirements, this dissertation sets up three kinds of planar object detection strategies through enhanced yang discrete-point grouping, including feature retrieving, boundary extraction and indirect acquisition. The related research results are applied in two aerial image projects, round oil drum detection and harbor warship detection.At the end of this dissertation, for enhancing the adaptability of Yin-yang discrete-point computing, a general method of feature discrete-point computing is induced from a large number of applications of the yin-yang discrete-point computing, and is used to address the problem of line segmentation in natural handwritten Chinese documents, which propose a sub-columns projection method with feedback for text-line extraction. Under the structured problem solving framework of the discrete-point computing, this method are guaranteed with correct results for most testing samples of natural handwritten Chinese document by carrying out four steps, including selecting, sampling and optimizing, grouping and feedback, and line-edge optimization of discrete-point in line segmentation. Our algorithm has a comparative result with the state-of-the art result using the HIT-MW database of Chinese handwritten text with much lower time complexity.In this dissertation, theoretical research is deeply carried out on detecting and extracting three kinds of shapes, including line shapes and planar shapes in complex scenes, and line shapes with complex structures. The method of yin-yang discrete-point computing is discussed in detail, and the method of feature discrete-point computing is proposed, which are all tested and improved in several real projects. This dissertation makes an attempt at and put efforts into solving the bottleneck problem faced by object detection in complex scenes and shortening the gap between computer vision and human vision.