Article Preview
TopIntroduction
Machine learning is a growing field, where machines are equipped with human features like sensing, hearing, seeing and smell. Computer vision is a field which facilitates the machine to recognize the object of interest in the same way as humans do. In the recent past there has been an extensive research in the area of computer vision. The recognition and retrieval methods proposed previously for 2D shapes do not completely provide convincing solutions. Objects in general have many features like shape, size, color, texture, mass and illumination. Among all the entities, shape is considered as the prominent feature for recognition and classification. Daliri and Vincent (2007). There are two criteria for selection of shape analysis, one is boundary-based approach, which is based on contour of the shape and the second is the holistic approach which is also known as area-based approach (p. 1782). Freeman (1974) wrote that an object represented by shape contour does not consider the internal marks or holes within the shape, which means the shape is represented by a single loop (p. 57). Bicego and Lovato (2012), Proposed a method where shapes are encoded as biological sequence, employing standard and well-established sequence alignment tools to device a similarity score (p. 1359).
Lovato et al. (2014), Proposed a new matrix called S-BLOSUM, which is used in determining the sequence alignment solution. This matrix learns the rate of matches and mismatches (p. 2335). Bribiesca and Wilson (1997), presented an approach for 2D shape object dissimilarity. The amount of dissimilarity is evaluated based on the transformation of one shape into another, by representing object to be compared are invariant under translation, rotation and Scaling. Bandera et al. (1999), proposed an algorithm, where curvature functions are used to represent the shape contour, further Fourier domain is used in decomposing the object as linear combination of a set of representative object, which is finally recognized by multilevel clustering (p. 49).
Guerra. C (1998), presented an approach, where the objects are represented as a tree structure by decomposing boundary in terms of convex/concave multi scale boundary. It requires reconfigurable mesh architecture (p. 83). Mcneill and Vijaykumar (2005), present a method where the points along the boundaries are placed at fixed intervals or radial angle. This is considered as the accurate and robust descriptor of shape. By taking distance and angle of basic descriptor a wide range of shapes can be classified accurately. (p. 1483).
Khalil and Bayoumi (2000), proposed a method where, the continuous wavelet transformation and neural networks in classifying of objects using translation, rotation and scaling transformations. (p. 863). Belongie et al. (2002), present a method to measure similarity between two shapes using shape context. The shape context at any reference point acquires the distribution of all the other points relative to it. This offers a globally discriminative characterization. So, the two points on two similar shapes will lead to same shape context (p. 509). Sun et al (2008), propose a method in which the boundary of the shape is extracted and represented as eigen-values of a covariance matrix for a region of consideration. These sequences are re-sampled and transformed using autocorrelation function. The method is invariant to translation, rotation and scaling (p. 1966). Haibin et al. (2007), presented an approach using inner-distance to represent shape descriptor, that is used to capture the object structure. The inner distance is the shortest path between the two landmark points within the shape. Texture information along the shortest path is used to improve shape classification (p. 286). Daliri and Torre (2007) proposed a method using shape context and dynamic programming to find the object best match (p. 1782).