Image Local Features Description Through Polynomial Approximation

Abstract

This work introduces a novel local patch descriptor that remains invariant under varying conditions of orientation, viewpoint, scale, and illumination. The proposed descriptor incorporate polynomials of various degrees to approximate the local patch within the image. Before feature detection and approximation, the image micro-texture is eliminated through a guided image filter with the potential to preserve the edges of the objects. The rotation invariance is achieved by aligning the local patch around the Harris corner through the dominant orientation shift algorithm. Weighted threshold histogram equalization (WTHE) is employed to make the descriptor insensitive to illumination changes. The correlation coefficient is used instead of Euclidean distance to improve the matching accuracy. The proposed descriptor has been extensively evaluated on the Oxford's affine covariant regions dataset, and absolute and transition tilt dataset. The experimental results show that our proposed descriptor can categorize the feature with more distinctiveness in comparison to state-of-the-art descriptors. INDEX TERMS Covariant, descriptor, handcrafted feature, patch, textures.

Application of Machine Learning for the Automation of the Quality Control of Noise Filtering Processes in Seismic Data Imaging

“…To overcome this issue, a new method for performing a nonlinear PCA was introduced in [11], which suggested computing principal components in high-dimensional feature space using integral operator kernel functions and was used for pattern recognition. In another context (i.e., feature matching) [12] highlighted the power of kernel (i.e., polynomial) approximation for a more efficient representation of the extracted patches (circular shape window around the key-points of a given image).…”