Efficient Deep Learning for Image Classification: Lighter Preprocessing and Fewer Parameters

Convolutional neural networks have achieved state-of-the-art performance in several computer vision tasks recently, learning high-level representations directly from RGB images. However, using deeper architectures has led to high computational costs, hindering deployment on devices with limited resources. Additionally, models are usually specialized in a single domain/task while an increasing amount of real-world applications need to deal with multiple domains simultaneously. The computational cost of storing and running multiple instances of those costly models can limit their utilization even more. This Ph.D. thesis aims to reduce the computational burden of deep learning, focusing on two main aspects: reducing data preprocessing cost and sharing parameters across multiple domains/tasks. These contributions have led to the creation of efficient models with high classification performance and reduced costs, allowing them to be deployed in a wider array of devices.