Binary gradient computation and implementation in reconfigurable computing environments

Anton Bondarchuk, Dmitriy Shashev, Stanislav Shidlovskiy
The article outlines a new approach to constructing a feature vector for implementation on computers with a parallel pipeline architecture. The feature vector consists of the calculated characteristics of the gradient of a binary image (binary gradient) by analogy with the operation of the HOG algorithm. The proposed algorithm detects features of the contour pixels of objects in a binary image, which are further used for pattern recognition. After using the newly generated feature vectors for training a support vector machine (SVM) classifier, the speed of processing and classifying objects of interest on an image with a size of 1280 × 720 pixels increased by 3.5 times, in comparison with using the classical HOG descriptor.