A NEW TOMOGRAPHIC BASED KEYPOINT DESCRIPTOR USING HEURISTIC GENETIC ALGORITHM

Keypoint descriptor is a fundamental component in many computer vision applications. Considering both computational complexity and discriminative power, SURF descriptor among non-binary and BRISK among binary descriptors are the prominent techniques in the field. Although, these descriptors have shown remarkable performance, but they are still suffering weaknesses such as lack of robustness against image transformations and distortions, especially blur, JPEG compression and lightening variation. To address this matter, a new and robust keypoint descriptor is proposed in this research which is adapted from Tomographic-Image-Reconstruction technique. Convolution of associated image patch and predefined Gaussian smoothed sensitivity maps yield a matrix whose entities demonstrate the average intensity of the pixels at the convolved pixels in the image patch. The initial descriptor vector is built by calculating the absolute differences of all possible pairs of matrix. Then, the most discriminative features of this initial descriptor vector are detected by Heuristic Genetic Algorithm (GA). The Experimental result showed that proposed keypoint descriptor outperformed some existing techniques especially in blur, JPEG compression and illumination variation while it has reasonable performance in other image transformations.