Advanced Learning-based Visual Content Analysis for Computer Vision

Abstract

This thesis tackles the challenges in deep neural networks by introducing advanced learning-based methods for visual content analysis tasks for computer vision.

The quest for improved deep learning solutions is ongoing. Computer vision applications use techniques such as classification, object detection, segmentation and counting both widely and extensively. However, these applications have challenges such as domain variability, poor generalization, inadequate localization, and incomplete and weakly annotated datasets which need to be addressed to obtain better application accuracy. Hence, there is a need for advanced solutions which address these challenges. In the thesis, which we have divided into two parts, we tackle these challenges within both fully supervised and weakly supervised settings, contextualizing them in various applications.

In the first part of our thesis, we develop sophisticated deep learning techniques for computer vision applicable to a fully supervised context. We present methods aimed at enhancing model generalization and refining object localization. We have studied the use these advanced deep learning methods in classification and object detection to demonstrate the effectiveness of our algorithms.

In the second part of our thesis, we extend our fully supervised approaches to handle incomplete datasets where the major challenges are domain gap and weak annotations, and apply it to object detection, person re-identification, and cell counting.