Research Of Affine Invariant Feature Extraction Algorithm Based On Shape Region Segmentation

With the development of modern technology, more and more images have been used to transport information. So now researchers focus on computer vision such as object recognition and image retrieval. For object recognition and image retrieval, object shape features are powerful because shape is usually strongly linked to object functionality and identity. In recent years, a wide range of shape-based feature extraction and similarity measures methods are proposed in literatures, moreover various applications are likely to use shape features.For shape-based feature extraction and similarity measure, geometric invariance and deformation resistibility are important issues. Nowadays, many approaches have been proposed which are divided into two classes:region-based and contour-based. Geometric moment, Hu invariant moment, Zernike moment, generic Fourier descriptors and so on are all region-based approaches, and Fourier descriptors, chain codes, curvature scale space and so on are all contour-based approaches. Especially, Fourier descriptors are classical approaches and realized easily.In this paper we proposed a novel shape signature for Fourier descriptors. The shape signature is called contour split-based area function. We extract the closed contour of the shape, and resample the contour with the method of iso-area normalization. Different from the traditional methods like centroid distance and farthest point distance which use distance as the functions, the proposed approach use area as the shape signature function. The contour of shape is split by equal number of points, so we can get the area of each region surrounded by the chord of linking start point and split point and the corresponding arc. All points on the contour can build a matrix to save the areas, and this matrix is the proposed shape signature. Assuming that the rows of the matrix represent the number of the split regions and the columns represent the number of points on the contour. Then we apply Fourier transform on every row of the matrix, and only select the first20normalized Fourier coefficients to build a new matrix. This matrix is the Fourier descriptors feature matrix. According to the proposed approach, we define a new approach of similarity measure based on Euclidean distance.The proposed approach has the qualities of translation, scale, and rotation invariance, and the great advantage of it is invariant to affine transform, because of the method of resample and using area as the shape signature function. In the paper we not only prove the affine invariance through mathematical deduction, but also do experiment to show the correctness of the conclusion. At last, we select the MPEG-7CE-Shape-1part B as the image database for retrieval. We select another six kinds of Fourier descriptors which are centroid distance, complex coordinates, cumulative angle function, area function, farthest point distance and multi-level chord length function descriptor to compare with the proposed approach. At the same time, we select two non-Fourier descriptors which are Zernike moments and curvature scale space to compare with the proposed approach.The result of the comparative experiment shows that the performance of proposed approach is much better than other compared approaches. And the bull’s eye of the proposed approach is79.60%which is more than ten percentage points higher than centroid distance which is proved as the best Fourier descriptors by D.S. Zhang. The novel approach for shape signature proposed based on contour split and using the split region areas as function describes shape not only with reference location of points on contour, but also with shape regions referenced, so it can represent shape more completely than other shape signature approaches. The proposed approach is affine invariance, and the experiments on image database show the approach’s advantage.