Recognition of objects is one of the basic tasks in computer vision applications. Recognition is usually based on gray levels or colors, and shape characteristics of target objects. The goal of object recognition is to find a description which contains sufficient information to distinguish between different target objects. The design of potential target objects is often known in advance, so the geometric used in creating the object (i.e., in manufacturing or in some natural formation process) can be directly applied to the recognition task. Consequently, the major difficulties might be merely in arranging the physical image acquisition. Plenty of methods exist for recognition of objects [ 14 , 12 , 17 ]. However, some of the methods, for example syntactical methods, are mainly suitable for regular or man-made object recognition. In this paper the main concern is on irregular objects (for example surface defects) which are hard to recognize even for a human observer.
There are several methods for the shape analysis of objects [ 14 , 12 , 17 ]. These methods can be divided into two categories, the area-based methods and the contour-based methods. The latter methods are of interest in this paper since we are mainly concerned on the shape of a contour. The contour-based methods include the following techniques. Simple descriptors, for example perimeter length, curvature, and bending energy, have been applied widely [ 12 , 17 ]. Moment-based techniques have been used in object recognition since 1962 [ 8 ]. Moments derived from the contour of an object were used by Dudani et al. [ 3 ] and Gupta and Srinath [ 7 ]. Zahn and Roskies [ 18 ] used the Fourier coefficients of a contour as shape descriptors. The chord distribution of a contour was proposed by Smith and Jain [ 16 ]. A scale-space technique to form a description for plane curves was proposed by Mokhtarian and Mackworth [ 13 ]. Evans et al. [ 4 ] proposed pairwise geometric histograms as shape descriptors. The chain code histogram was proposed by Iivarinen and Visa [ 10 ].
In Section 2 three different shape coding techniques are shortly introduced. The chain code histogram (CCH) is a statistical measure for the directionality of the contour of an object [ 10 ]. It is calculated from the chain code presentation of the contour. The pairwise geometric histogram (PGH) was introduced by Evans et al. [ 4 ]. Lately Ashbrook et al. developed the algorithm further [ 2 ]. The PGH is a shape descriptor which is applied to polygonal objects. The last technique uses simple shape descriptors [ 12 , 17 ]. The combination of five such descriptors seem to provide good object recognition capabilities [ 15 ]. In Section 3 the discriminatory powers of these three shape coding techniques are demonstrated on real irregular objects. Further discussions on the properties of these techniques are provided in Section 4. Conclusions are drawn in Section 5.
Jukka Iivarinen