Optimization for Deep Learning Systems Applied to Computer Vision

Abstract

Since the DL revolution and especially over the last years (2010-2022), DNNs have become an essentialpart of the CV field, and they are present in all its sub-fields (video-surveillance, industrialmanufacturing, autonomous driving, ...) and in almost every new state-of-the-art application that isdeveloped. However, DNNs are very complex and the architecture needs to be carefully selected andadapted in order to maximize its efficiency. In many cases, networks are not specifically designed for theconsidered use case, they are simply recycled from other applications and slightly adapted, without takinginto account the particularities of the use case or the interaction with the rest of the system components,which usually results in a performance drop.This research work aims at providing knowledge and tools for the optimization of systems based on DeepLearning applied to different real use cases within the field of Computer Vision, in order to maximizetheir effectiveness and efficiency.