Multi-scale Multi-block Covariance Descriptor with Feature Selection

Multi-scale Multi-block Covariance Descriptor with Feature Selection Abdelmalik Moujahid1, Fadi Dornaika1 1University of the Basque Country UPV/EHU, San Sebasti´an, Spain Published in: Neural Computing and Applications, Springer-Verlag, 2020. DOI: 10.1007/s00521019-04135-7. Published online 15 March 2019. Abstract This paper investigates a compact face texture representation able to cover the most discriminant features of facial images. The compactness is achieved by the proposed Pyramid Multi-Level (PML) covariance texture descriptor and the feature selection process that is applied on the raw extracted features. In fact, we introduce a framework based mainly on two new aspects. Firstly, we consider an extension of the original covariance descriptor that relies on de-noised covariance matrices obtained using texture descriptors such as local binary pattern and quaternionic local ranking binary pattern images. Secondly, we exploit the resulting covariance descriptor using a PML face representation which allows a multi-level multi-scale feature extraction. Experiments conducted on four public face datasets show the efficacy of the proposed face descriptor and the associated selection schemes. Main Contributions •Proposes the Pyramid Multi-Level (PML) covariance descriptor, enabling multi-scale, multi-block representation of face images using second-order statistics from diverse texture features. •Extends traditional covariance descriptors through explicit pyramid constructions and de-noising based on random matrix theory, yielding adaptive and robust face representations. •Introduces attribute and block selection strategies using supervised Fisher scoring to extract the most discriminative features and blocks, reducing redundancy and improving classification accuracy. •Provides a thorough evaluation on four benchmark face datasets (Georgia Tech, Yale, ORL, FEI), demonstrating that the proposed PML covariance descriptor outperforms PML-LBP and PML-HOG and rivals advanced classification approaches. 1 •Supplies fine-grained visual analysis (block relevance maps) and extensive benchmarking against other state-of-the-art methods, showing the advantages of multi-block and multi-scale feature fusion. Impact of the Paper The study advances face image representation and texture analysis by combining multiscale, multi-block partitioning with robust covariance-based feature extraction and targeted feature selection. This approach achieves highly compact, discriminative descriptors that encode rich spatial and statistical information from facial images, enhancing both local and global feature capture. The denoising and multi-level fusion strategies allow the descriptor to adapt to varying image qualities, face poses, and illumination conditions, which are critical challenges in practical face recognition systems. The methodology’s strengths include: •Flexible and generalizable representations: The descriptor can incorporate and combine multiple types of features—such as LBP, QLRBP, gradients, and color channels—making it adaptable for various face analysis tasks and robust to environmental variations. •Efficient dimensionality reduction: The use of Fisher scoring and block analysis enables efficient selection of the most informative regions and attributes, leading to lower computational costs and improved performance, as demonstrated on large and diverse datasets. •Advancement of interpretable AI: Visual maps of block relevance not only support better accuracy but also offer insights into which face regions and features drive recognition, fostering transparency in computer vision models. •Benchmark-leading results: The approach outperforms traditional descriptors and matches or exceeds more complex classifiers on benchmark tests, making the strategy viable for real-world applications beyond just research settings. Overall, this paper sets a precedent for combining multi-scale texture analysis with principled feature selection, contributing directly to the development of next-generation face recognition technologies that are both robust and resource-efficient. Reference: Abdelmalik Moujahid, Fadi Dornaika. “Multi-scale multi-block covariance descriptor with feature selection.” Neural Computing and Applications, Springer-Verlag, 2020. DOI: 10.1007/s00521019-04135-7 2