Object detection is one of the most fundamental tasks within the field of computer vision. In recent decades, deep learning-based approaches to object detection have attracted significant attention. Despite the increasing scientific evolution of object detection algorithms, solutions given by pre vailing deep learning-based methods are still limited due to challenges such as data limitation. Challenges due to data include the enormous human labor cost to label positive data for large number of object classes leading to the problems such as data imbalance and lack of data for domain specific objects. This thesis aims to develop deep learning-based techniques to advance the performance of object detection by tackling the data challenges. Specifically, an innovative framework is firstly developed for directly detecting seat belts in surveillance images, bypassing the need for detecting auxiliary objects such as cars and windshields by using a Part-to-whole Attention mechanism and a Gate Bi-directional LSTM on the patches diagonally sampled from region proposals potentially containing seat belts. Secondly, a novel regression loss function named Adaptive Enhanced Corner Intersection over Union (AEC-IoU) loss is devised to alleviate the imbalance problem by a soft-weighting strategy and enhance corner alignment to accelerate convergence speed with a simple weighting factor approach that is able to reshape the existing IoU-based losses according to a geo metric relationship of bounding boxes. Thirdly, a weakly supervised object detection model is developed by incorporating rich contextual information, thereby minimizing the reliance on intensive human annotation for training samples. Two context proposal mining strategies and a Symmetry Context Module (SCM) are proposed to better capture the diverse discriminative information for objects of interest, improving the localization and detection performance of the detection framework trained by weakly labeled data.
- Object detection
- deep learning
Deep learning-based object detection tackling data challenges
Gu, X. (Author). Jul 2025
Student thesis: PhD Thesis