The Histogram of Enhanced Gradients (HEG) - A Fast Descriptor for Noisy Face Recognition

Descriptors serve as algorithms responsible for the representation and processing of multimedia files.One of the main challenges in these algorithms involves establishing a tradeoff among conflicting performance requirements such runtime, on the one hand, and standard metrics such as accuracy, precision, recall and F-score, on the other hand.To address this challenge, a novel descriptor named Histogram of Enhanced Gradients (HEG) is introduced for noisy face recognition.The methodology behind HEG involves enhancing local gradients prior to feature extraction, corroborated by the Histogram of Oriented Gradients (HOG) descriptor and adaptive filtering.Initially, facial images are divided into blocks, and features are extracted from each block using magnitude and orientation maps to discriminate between edges, details, and flat regions.Then, these features undergo denoising with an adaptive anisotropic diffusion filter, individually customized for each of these three types.Subsequently, the enhanced histograms from the blocks are concatenated to create a comprehensive feature vector representing the original noisy face image.Finally, the HEG descriptor is integrated within a supervised machine learning scheme with a Support Vector Machine as the classifier.The proposed descriptor is evaluated not only in terms of runtime and the standard metrics cited above, but also on the basis of six other similarity metrics, across three online datasets.Experimental results, conducted under different noise levels, clearly demonstrate that the HEG descriptor outperforms nine stateof-the-art descriptors on all three datasets yielding significant enhancements in runtime efficiency, with speed improvements ranging from 1.64 to 29.56 times, and notable refinements in F-score, ranging from 1.03 to 2.39 times.These results highlight the effectiveness of the HEG descriptor in capturing facial features from multimedia noisy files.