Local Extrema Min-Max Pattern: A novel descriptor for extracting compact and discrete features for hand gesture recognition

Abstract

In this paper, a handcrafted feature descriptor namely Local Extrema Min–Max Pattern (LEMMP) is introduced for static hand gesture recognition (HGR). LEMMP thoroughly characterizes the discriminative information between a specific coefficient and its nearby neighbors within a local window. The proposed approach extracts 2nd-order local information by encoding the most informative directions contained within multiple discrete spatial associations from each neighborhood pixel using a two-structure encoding method, which helps in detecting the gray level variations that may occur in different directions. LEMMP encodes the structure of hand gestures utilizing texture information in an easy-to-understand and compact coding scheme, resulting in improved accuracy with less memory and time as compared to existing approaches. Furthermore, the features extracted by the proposed LEMMP are classified using SVM classifier. The proposed LEMMP is tested on nine benchmark HGR datasets. The experimental results and the visual analysis demonstrate that the proposed LEMMP outperforms the existing state-of-the-art approaches with an accuracy of 52%(ASL Static), 92% (MUGD Set1), 98% (MUGD Set2), 75%(MUGD Set3), 64%(MUGD Set4), 71%(MUGD Set5), 98%(ASL Digit), 85% (NUS-I Dataset), 80% (NUS-II Dataset), 99% (ASLFS A), 99% (ASLFS B), 99%(ASLFS C), 96%(ASLFS D), 99%(ASLFS E), 39% Bengali Sign Language, 70% (HG-14 Dataset) and 42% (OU Hands) respectively.