Invariant Feature Extraction Using Central Projection

The extraction of image invariant feature is a hot topic in the field of image processing, computer vision, pattern recognition etc, and has extensive applications. The feature extraction approaches can be mainly classified into two main categories: contour-based methods and region-based methods. The contour-based methods extract features from the contour only, and its implementation is easy and it takes small calculation complexity. But these methods are usually not able to apply to the objects with several separable components (like some Chinese characters). On the other hand, the region-based techniques take all pixels within a shape region into account to obtain the shape representation. The derived features can reflect all the image characteristics. However, these methods are usually at the expense of high computational demands and sensitive to noise in the background of the image.In this paper, we propose a novel method by combining the contour-based methods with region-based methods for invariant feature extraction. A closed contour can be derived from the object by central projection transformation (CPT), and then different invariant features are extracted from this closed contour. By performing projection along the line with specific angles, this contour is able to reflect the region information of the image. Meanwhile, the global information of the image is presented as the angle changes. The proposed method takes less amount of calculation complexity and can extract effective invariant features.Main results of this thesis are as follows:(1) Propose a novel Fourier descriptors based on CPT. The original CPT method is considered to extracts rotation invariant features only. The proposed descriptors are invariant to shift, rotation and scaling transformation by combining Fourier descriptors with CPT method. To confirm the effectiveness of the proposed descriptors,52 English letters and 70 similar Chinese characters are used in the experiments.(2) The affine invariance of the closed curve derived by CPT has been proved, namely the affine transformation between the closed curves is the same affine transformation between the transformed images. We also provide an algorithm to extract affine invariant features making use of the derived curve. The curve is first parameterized to establish a one-to-one correspondence between points under affine transformation. The affine invariant features are derived from the closed curve after stationary wavelet transform. The computational complexity of the algorithm is also discussed. In the following experiments, 105 trademark images,81 Chinese characters and 26 capital letters are used to evaluate the proposed descriptors. Compared to MSA and AMIs, the proposed method not only can efficiently extract invariant features with lower computation complexity but also is robust to noise.