SmartDehaze: Haze Recognition and Removal System for Smart Vehicles

Abstract

Haze is one of the most common atmospheric phenomena. In hazy weather, water vapors, dust, and other airborne particles obscure the clarity of the scenery objects, severely degrading the quality of the outdoor images taken by camera. The reduced visibility impacts object detection and recognition, lowers the reliability of outdoor surveillance systems, and obscures the satellite images. In consumer-level photography, haze changes the colors and reduces the contrast of the photos. The degradation cannot be avoided by a better camera or lens, because it happens in the atmosphere before reaching the apparatus. The objective of this thesis is to develop a practical "smart-dehaze" system for the smart vehicles, that can judge and restore visibility, if needed.

The current dehazing techniques are not robust enough to perform well in adverse circumstances, also their runtime increase nonlinearly with image-resolution, which makes them unsuitable for real-time applications. To achieve real-time halo-free dehazing, a novel transmission refinement scheme called "reliability-guided fusion" has been developed. This method handles the challenging bright-regions well, and generates a high-quality refined transmission map with edge-preserving and texture-smoothing properties. Quantitative and qualitative comparisons show that the proposed dehazing scheme outperforms the state-of-the art techniques in terms of speed, quality, and reliability (handling difficult situations).

Image dehazing on a clear-day is an unnecessary burden that should be avoided. Through the use of edge- and haze-density features, and temporal coherence based decisions, we achieved very accurate and fast scene classification for hazy and clear weather conditions. The proposed "smart-dehaze" system is a cascade of "scene classification" and "dehazing" modules that can judge and restore visibility, if necessary. By combining smart-dehaze with (a clear day) pedestrian detector, we were able to accurately detect pedestrian equally well in clear and foggy weather conditions.

We have developed two datasets to quantify the effectiveness of the smart-dehaze system, and its cascade with object detection techniques. The first dataset consists of hazy and non-hazy images that were captured from a vehicle mounted camera. This dataset is used for training and testing of the proposed "scene classification" module. The second dataset consists of hazy images with annotated pedestrians. These annotations serve as ground-truth to judge detection accuracy of state-of-the-art pedestrian detector in hazy and dehazed conditions. Dehazing performance of several haze-removal algorithms is reported in terms of pedestrian detection accuracy boost they can provide on the hazy pedestrian dataset. Detailed evaluations show that, the proposed "smart-dehaze" system enables the pedestrian detector to approach its peak classification accuracy on both hazy- and clear-day pedestrian datasets, by dehazing only when necessary.