Segmentation of Structured Objects in Image
Rajeshwari Rasal · 2015
Detection of foreground structured objects in the images is an essential task in many image processing applications. This paper presents a region merging and region growing approach for automatic detection of the foreground objects in the image. The proposed approach identifies objects in the given image based on general properties of the objects without depending on the prior knowledge about specific objects. The region contrast information is used to separate the regions of the structured objects from the background regions. The perceptual organization laws are used in the region merging process to group the various regions i.e. parts of the object. The system is adaptive to the image content. The results of the experiments show that the proposed scheme can efficiently extract object boundary from the background. Image segmentation is the task of dividing an image into coherent regions so that each region corresponds to an object or area of interest. The detection of structured foreground objects in an image is useful in many applications like image classification, image retrieval, content-aware image resizing etc. The structured objects are more difficult to identify as they are composed of multiple parts. Different parts of object may have distinct surface characteristics (e.g. colors, texture etc.). The proposed approach tries to detect the structured foreground objects in the image using region contrast information and perceptual organization of object parts. The approach is based on the general properties of real-world objects (e.g. similarity, proximity etc.) and hence does not depend on specific properties of objects. The structured objects usually have high contrast to their background. Hence region contrast information is used to separate the regions that belong to unstructured and structured objects. Perceptual organization is the basic capability of human visual system to identify relevant groupings and structures from an image without any prior knowledge of its content. The Gestalt Laws are based on human visual perception of objects. In the proposed approach Law of similarity, Law of Symmetry, Law of Alignment & Law of Proximity are used to group the regions together to form a region that constitutes a structured object in an image. The proposed approach applies Gestalt laws to image regions and merge the regions together to identify one single region that represents a structured object. The accuracy of region merging is measured by using boundary energy function. segmentation approach each pixel of the image is equivalent to a node in the graph and edges represent adjacent pixels. Weights on each edge are the dissimilarity between pixels. The boundaries between regions are defined by measuring the dissimilarity between the neighboring pixels. In the Ncut method the nodes are arranged into groups so that within the group the similarity is high and between the groups the similarity is low. In the region-based approach, each pixel is assigned to a particular region. In region growing method of image segmentation a seed region is selected at start and a new region is identified by merging as many neighboring regions with the seed region. The boundary detection of objects was also implemented as a supervised learning problem (12). A large data set of human- labeled boundaries in natural images is used to train a boundary model. The model can then identify boundary pixels based on a set of low-level cues such as brightness, color and texture extracted from local image patches. In the multi-class image segmentation technique (9) a number of classes (e.g., road, sky, water, etc.) are defined for labeling every pixel in an image. In this method the image is first segmented into multiple coherent initial regions and then each region is assigned to one of predefined classes. Gestalt grouping laws can also be used for segmenting an image. They used boundary energy functions to identify the regions that can be grouped together. The energy function includes information about the region. The energy function measures the accuracy of grouping the regions.