DefinePK

Abstract

Many computer vision applications rely on matching key points between images. Over recent decades, advancements in key-point detection algorithms have significantly improved both robustness and speed. However, there is an ongoing need for more compact descriptors and faster methods with higher classification accuracy. This work addresses this need by introducing a novel algorithm that formulates both a key-point detector and descriptor based on prominent image features. The proposed detector focuses on identifying intensity-based corners and edges within grayscale images. This process involves detecting connected regions by analyzing pixel intensity ranges. Once bright and dark regions are identified, pixel intensities are sorted accordingly. Symmetric sampling is then applied after cascade matching, utilizing 128-bit descriptors. Isotropic and anisotropic filtering techniques are applied to the maximum filter response of the grayscale image. To normalize the descriptors, L2 normalization is performed on the RGB query image. The resulting feature vectors are spatially organized, and Principal Component Analysis (PCA) is applied to reduce their dimensionality. To improve search efficiency, indexing and searching are performed based on a visual words representation of the database of visual features. The proposed method was evaluated using two datasets, Caltech-256 and Corel-1000, and compared to the standard HOG detector and descriptor. Experimental results show significant improvements in both average precision and average recall for the proposed method.