ADAM-Dehaze: Adaptive Density-Aware Multi-Stage Dehazing for Improved Object Detection in Foggy Conditions
Alhindaassi, Fatmah Ali Abdulla Mohamed
Alhindaassi, Fatmah Ali Abdulla Mohamed
Supervisor
Department
Machine Learning
Embargo End Date
2025-05-30
Type
Thesis
Date
2025
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
Adverse weather conditions, particularly fog, severely compromise the performance of standard object detection systems in urban environments and critical applications such as crime scene investigations and autonomous driving. Despite advances in computer vision, current methodologies often falter under varying fog/haze intensities, hindering timely and accurate scene analysis. This thesis addresses the challenge by hypothesizing that integrating the classification of haze intensity with specialized dehazing networks can significantly improve object detection accuracy in foggy conditions, with a particular focus on scenarios of high-intensity fog. The proposed end-to-end framework begins by categorizing the fog intensity into three distinct levels: light (0.03), medium (0.06), and complex (0.09), using a dedicated classifier on a dataset of 25,000 annotated images. For each haze level, an intensity-specific dehazing network is applied to restore image clarity before a robust object detection module processes the dehazed images. To ensure effective downstream performance, the framework incorporates an evaluation step that compares the detection accuracy on the dehazed images with the foggy conditions of the baseline, validating the impact of the dehazing process. Special emphasis is placed on high-fog intensity (complex) conditions, representing the most challenging environment for vision systems. By tailoring dehazing strategies to the precise level of atmospheric interference, the framework optimizes both dehazing and detection tasks, ensuring enhanced robustness in diverse realworld scenarios. Comprehensive experimental evaluations demonstrate that this integrated approach improves object detection performance in challenging foggy conditions and offers significant practical benefits. Detection accuracy and reliability improvements are advantageous for safety-critical domains such as autonomous vehicle navigation and forensic analysis at crime scenes. Key contributions of this research include the novel integration of haze classification with custom dehazing networks and the validation of its effectiveness through extensive testing, which paved the way for more resilient vision systems in adverse weather conditions.
Citation
Fatmah Ali Abdulla Mohamed Alhindaassi, “ADAM-Dehaze: Adaptive Density-Aware Multi-Stage Dehazing for Improved Object Detection in Foggy Conditions,” Master of Science thesis, Machine Learning, MBZUAI, 2025.
Source
Conference
Keywords
Dehaze, Adaptive, Detection