Developing Efficient Convolutional Neural Networks Models for Image Classification

Abstract

newline Deep architectures are gaining widespread recognition owing to their impressive performance newlinein addressing intricate problems. Among these architectures, Convolutional newlineNeural Networks (CNNs) have demonstrated exceptional efficacy across a broad spectrum newlineof image-related tasks such as image classification, object detection, and image segmentation. newlineUnlike prior research that relied on expert-designed feature representations, newlineCNNs endeavor to automatically learn representation hierarchies from the available data. newlineTheir multi-layered structure proves advantageous in capturing the hierarchical nature of newlinethe input data, as lower layers detect simple features, which are then utilized by higher newlinelayers to extract abstract representations. newlineNotwithstanding the remarkable representational capabilities of CNNs, training these newlinemodels is computationally demanding. The CNN s architecture can significantly vary in newlineterms of the number and types of layers, the quantity of fully connected layers, and the newlinestride and padding associated with the kernels, along with other hyperparameters defining newlinethe architecture. Moreover, a shortage of labeled training data in many applications newlineposes inherent challenges for the gradient-based optimization procedures used in parameter newlinelearning for these models. Consequently, extensive and ongoing efforts are being made newlineto discover algorithms that can automatically develop CNN architectures. newlineThis thesis addresses the formidable task of devising Convolutional Neural Network newline(CNN) architectures specifically tailored for image classification. It commences with a newlinecomprehensive investigation into the functioning and various types of CNN architectures, newlineas well as the challenges associated with their development. It details the different kinds newlineof approaches that have been utilized in the designing of CNN architectures. Furthermore, newlinethe dissertation offers a comprehensive exploration encompassing contemporary newlinealgorithms and methodologies pertaining to Neural Architecture Search (NAS) within the newlinedom