Research On Key Technologies Of Salient Region Detection And Indoor/Outdoor Scene Classification

Image Understanding mainly includes two aspects:object detection and scene classification. In the real world, the existing of the objects is not mutually isolated, and has some relationship with its surrounding. Therefore scene understanding provides prior information to guide the object detection process. Scene understanding can be divided into two types:First, focusing on the psychology and physiology. Its main research are scene quickly perception both in psychological and physiological (e.g. salient region detection); Second, focusing on computation models for scene classification. Its main research is scene analysis using simple statistical models (e.g. Indoor/outdoor scene classification). Traditional saliency algorithms are based on human visual models and can only generate the salient points, limiting its further application. So how to detect the salient region from an image becomes a new target. In addition, as lighting conditions, scale changes, accuracy scene classification could be a difficult problem. So how to extract features that are robust to these conditions for scene understanding becomes another hot research in image understanding.This thesis first surveys the existing work on salient region detection and scene classification. Based on this survey, this thesis presents two novel salient region detection algorithms, and one novel feature for indoor and outdoor image classification. The contributions of the thesis mainly include:1. As most of the algorithms are based on the center-surround differences of intensity, edge and color information, and their results usually emphasize the high-contrast edges instead of regions/objects. In this thesis, we distinguish salient regions from their backgrounds through texture, the contrast of object-covering properties. We propose a new salient region detection algorithm assuming salient regions and their backgrounds have different amplitudes of oscillation. Through interpolations of the intensity of local extreme, we obtain different oscillation amplitudes for object and its background respectively, with well-defined boundaries retained. Then the saliency map is computed through the difference of the amplitudes. Experiments show that saliency generated from our method can uniformly highlight the whole salient region.2. The frequency-tuned method using the mean value of the image as the redundant information. However, for most images, using the mean will lower the difference between salient regions and the backgrounds, which makes it difficult to discern between backgrounds and salient regions. In this thesis, we take the background information as redundant information, and remove the background information using color interpolation. Experiments show that the proposed salient region detection method offers three advantages over existing methods: high contrast between salient regions and backgrounds, uniformly highlighted salient regions and high precision and good recall rates.3. Traditional scene classification algorithm, using only low-level features (such as color, edge and texture), without considering spatial or semantic information. In order to extract mid-level cues, super-pixels are used as process units, which will add a natural spatial support to compute features. Besides, the super-pixels of images are usually the semantic objects or part of them, which make them an ideal candidate for image classification. Experiment shows that mid-level features can provide meaningful information for image indoor/outdoor classification.