Evaluation of shape classification techniques based on the signature of the blob

Abstract

In this paper, we report a study of different preprocessing and classification techniques that can be applied to shape classification using the signature of the blob, or its FFT, as the main feature. Eight well-known classification methods were tested and compared.

The results obtained show that, for shapes with a small to medium amount of distortion, all the methods obtained an almost 100% success probability. However, as distortion increased, those not based on the FFT performed better than the other algorithms, at the expense of a small increase in computational time.

The samples employed for training and testing purposes were not hand-selected, but were generated by an application developed as part of this study. This application simulates the main distortions that can be produced by a real camera, including shifts, scalings, rotations, affine transformations and noise. We demonstrate that the use of these synthetic images for the training process, instead of manually selected ones, had proven to perform well with real images.

A study of the false positive problem is also included, showing that, with the use of SVMs and careful selection of the training set, a large number of false positives can be discarded in the detection step.

Introduction

Shape classification can be seen as the task of identifying the shape of an object according to a set of previously defined reference shapes, including, if necessary, false positives as another class. Many object recognition problems are more easily solved in terms of shape recognition than with other features, such as color or texture. In recent years, the problem has been given considerable attention, with potential applications for image processing problems such as handwriting recognition, industrial inspection or computer aided diagnosis.

The problems that must be solved are, firstly, the choice of a shape descriptor that encompasses all the relevant information concerning the shape as accurately as possible, and secondly, the design of a classifier that, using the previously defined descriptor, compares and classifies each shape into one of the predefined reference shapes. Much effort has been expended on the first step, and many approaches have been described in the literature. Nevertheless, it is worth noting that the choice of the descriptors is driven by the kind of shapes to classify.

Some of the most popular descriptors reported in the literature are the well-known Fourier Descriptors (FD) [1]. The FD are used, for instance, in [2] for the identification of American Sign Language hand shapes, or more recently, in [3], [4], for the recognition of paper and metal defects, or human silhouettes. In [5], FD were extended to become the Generic Fourier Descriptors (GFD), for the retrieval and identification of pictographs.

Another well-known shape descriptor is the Contour Curvature [6], where a one-dimensional curve can be obtained from the curvature of the object contour. Some examples can be found in [7], for instance, for the classification of general objects, or in [8], for the recognition of diatoms, a kind of unicellular algae. In [9], the curvature of the contour was further tested with occluded objects. An extension of the curvature of the contour is the Area Matrix, which is more affine-invariant than the curvature itself, and was tested, for example, in [10] for the recognition of sign language hand shapes.

Another way to describe a shape is through the use of Zernike moments. Zernike moments can be found in a number of papers, such as [11], [12], [13], [14]. In these studies, the Zernike moments were used to classify different kinds of shapes, such as traffic signs, complex pictographs and handwritten texts. [15] reports a comparative study of Zernike moments and Fourier Descriptors.

Other shape descriptors in the literature include the Level Sets. In [16], the shape descriptor of known objects, such as handwritten text, was used for the segmentation of low quality images. In [17], [18], the Level Set Function was used for the identification of different shapes in medical images, such as Magnetic Resonance or Computed Tomography scans.

Although all the descriptors mentioned so far are capable of describing complex shapes, their main disadvantage is reduced performance in the presence of high geometric distortions, especially segmentation noise. In this paper, we focus on the use of the signature as the main descriptor of the shape of the object [1]. Shape signature has not been widely reported in the literature, most probably because it cannot describe complex shapes. However, the signature has proven to be a very robust and efficient descriptor of simple shapes in the presence of noise and geometric distortions, as demonstrated here.

The research reported here represents the continuation of previous research described in [19], [20]. Although these studies were exclusively oriented to traffic sign systems, part of our efforts has been devoted to the design of a more generic algorithm, that is, to be able to process and recognize a wider range of predefined shapes. Thus, we conducted a study comparing the performance of several classifiers which has enabled us to improve the performance of the algorithm in terms of classification success. All the results were obtained with the use and processing of synthetic images generated automatically with an application which was also developed as part of this research, although the algorithm was also tested with real images. Lastly, we investigated the false positive detection problem, which enabled the algorithm to eliminate those objects that did not correspond to any of the predefined shapes, thus reducing the computational complexity of the entire system.