LBP Inspired Efficient Deep Convolutional Neural Networks for Visual Representation Learning

Over the past decades, deep neural networks (DNNs) have revolutionized computer vision by excelling in various tasks. Recent efforts have mainly emphasized accuracy, leading to the development of large and complex models. However, as edge devices like mobile phones and embedded systems become more prevalent, the need for efficient computer vision models has grown. Feature representation quality is pivotal in computer vision, directly impacting model performance. This thesis addresses the challenge by enhancing traditional local binary pattern (LBP) descriptors for more discriminative features and creating compact DNN modules with reduced computational demands and model size, combining both approaches to meet the efficiency requirements of modern computer vision models.