Directional Selectivity Model And Its Application In Image Processing

Image understanding is one of the key problems in the field of computer vision. As far as image understanding is concerned, an important step is extracting significant features which constitute the semantics of the scene and representing them in a form which can be easily utilized by higher level tasks. A traditional method of feature extraction is searching for edges first and the forming lines or combining contours. These features provide high-level semantic understanding with basis.To some extent, the traditional methods of features extraction have made big progress. Many algorithms have been proposed for edge detection, line detection as well as contour detection. However, these algorithm are es-sentially dependent on edge detection, and therefore have several significant problems. The problems include high dependency on the results of edge detection, and too many parameters which need to be adjusted manually. Generally speaking, these algorithms all attempt to define the problem of image understanding from a computational aspect, but they usually neglect the essence of the problem, i.e. the physiological mechanism how biological visual systems accomplish edge-line/contour detection. As a result, it is very difficult for these algorithm to solve these problems fundamentally, and even more difficult to solve the higher-level image understanding tasks based on the results of these algorithms.As for human eyes, the most salient features in a visual scene are lines, curves and more complex shapes formed by different colors. Essentially, all the lines and curves we see can be considered as the combinations of short line segments of different orientations. Accordingly, this thesis takes orien-tations (short line segments) as one of the most significant kinds of features for image understanding. Obviously, orientations, as primary features, are higher than low level features such as gray scales or colors of pixels, and lower than high level semantic features such as contours. As a multidisci-plinary research on neuroscience-based artificial intelligence and cognitive science, this thesis attempts to design a bio-inspired method with the neural mechanism how visual systems detect orientations. This thesis aims to solve the basic problem of feature extraction and representation in order to provide more complex tasks with physiological foundations.In visual neuroscience, simple cells’orientation selectivity has been a hot topic for a long time. Nobel Prize laureates Hubel and Wiesel put forward a model of a simple cell’s receptive field. This model accounts for the orien-tation selectivity of a simple cell to bar stimuli using geometric constraints. On one hand, this model has a simple form; on the other hand, it has several defects and is therefore being challenged. However, no strict physiological evidence has directly proven or denied this classical model yet till now.On the basis of Hubel-Wiesel neural model, this thesis proposes a double-layer network model of orientation computation. Compared with Hubel-Wiesel model, the proposed model has fewer limitations on the bottom-layer neurons, and less requirements on the stimuli. The proposed model imple-ments the computational detail of each layer, and is therefore more flexible. The numeric experiments indicate that the proposed model can simulate the orientation selectivity of simple cells well.With the proposed model of orientation computation, this thesis further puts forward an algorithm for detecting orientations in digital images. The experiments performed on synthesized and natural images indicates that this algorithm can extract satisfactory orientation maps from complex images. Compared with the edge/line maps obtained with traditional algorithms, the orientation maps highlight salient semantic features, while suppress trivial distractions. Moreover, they are closer to the real contour maps of objects, and therefore improves high level tasks such as segmentation and recogni-tion. Furthermore, the proposed algorithm is much less dependent on pa-rameter adjustment, and can be realized with highly parallel computation.As an application, this thesis skillfully uses the orientation detection al-gorithm to explain several famous geometric illusions. This thesis explores the local mechanisms of visual illusions with a qualitatively computational approach, and generates illusive phenomenon which is similar to human eyes’ observation and understanding. This thesis then explains a series of illusions with this approach. As a further application, this thesis tentatively uses the orientation maps of monocular image to partially recover the three dimensional information of the scenes, and eventually makes some achieve-ments.