Haze removal algorithm based on single-images with chromatic properties

Highlights

• A haze removal technique is based on hazy state detection and simple priors (DRM).

• Hazy state detection serves to identify the fog in an image by the linear relationship between brightness and saturation in HSV color space.

• Based on that fog and highlight appear to share the same chromatic properties, DRM can approximate the fog map and object function is to refine haze-free image.

• Proposed method is 2.5% higher than other methods in road segmentation accuracy after haze removal.

Abstract

Dense fog conditions formed by particles and water droplets in the atmosphere impact the road segmentation accuracy of Advanced Driver Assistance Systems (ADAS) by degrading the contrast and color fidelity of images. Existing road segmentation methods are not suitable for dense haze conditions; darkness, under-estimation, and over-enhancement problems occur after dehazing. This paper proposes a single-image haze removal technique based on hazy state detection and simple priors. Hazy state detection serves to identify the fog in an image by the linear relationship between brightness and saturation in HSV color space. Simple priors both in the background and hazy layers are based on a dichromatic reflection model (DRM) to accommodate different concentrations of fog. We find that according to color consistency and physical properties, fog and highlight appear to share the same chromatic properties. Based on this observation, the DRM can approximate the fog map. The maximum a posteriori formulation can also be utilized to robustly refine the fog map and remove black noise. We employ the fog map to compute atmospheric light and predict transmission rate in order to obtain haze-free images. Our results show that this method outperforms existing methods in regards to both efficiency and dehazing ability.

Introduction

Safe, reliable vehicles critically depend on autonomous driving and intelligent transportation systems. Most ADAS perception systems contain LIDAR-based primary detection blocks and vision-based road segmentation modules. The vision-based system is vulnerable to inclement weather such as fog and rain, which can cause a severe increase in traffic accident rates. Fog or rain come with large amounts of particles and water droplets which absorb and scatter light in the atmosphere, causing blurring, poor contrast, and degrade fidelity in outdoor images. ADAS are substantially limited in harsh weather conditions, to this effect. There is urgent demand for efficient and robust algorithms to benefit computational photography and computer vision applications such as stereo matching [1], segmentation [2], and recognition [3].

Haze (e.g., fog) removal is a challenging endeavor because the concentration of the haze is related to depth, which is difficult to assess in a given image. Early researchers on this topic attempted the use of multiple images or additional information [4], [5], [6], [7]. Although these methods perform well in capturing the depth of haze across multiple images, they are not suitable for real-time application due to excessive input requirements.

Recently, single-image-based haze removal has attracted a great deal of research interest due to its potential for much broader application. Many single-image methods are based on priors or assumptions for effective haze removal. Haze can be removed efficiently by maximizing the local contrast of a hazy image, and the depth information is gathered by learning strategies. These contrast enhancement methods above also perform well but over-saturation always occurs in white regions and results in abnormal color in local areas, such as sky or white tall buildings. For haze removal based on deep learning, they perform excellently in removing haze with fast speed; however, the results are affected by training dataset and may fail in heterogeneous fog conditions. In this study, we combined the DRM with a probabilistic framework based on single images (Fig. 1). This technique can significantly enhance scene contrast and visibility in images featuring dense fog. Instead of using the DCP or boundary constraint, we use the DRM [8] to initially estimate the fog map. Based on our observation that highlight and fog share the same chromatic properties, we assert that haze removal can be treated as a decomposition problem similar to highlight removal. The chromatic properties of the fog map can constrain the chromaticity of restored images and prevent over-saturation.

It is important to note, however, the slight differences between fog and highlight which introduce noise to the fog map and de-hazed images. In our previous work [11], we used a bilateral filter to remove noise and haze from images. However, noise sharply increases as image sizes increases; the bilateral filter no longer functions beyond a certain image size. In this study, we attempted to streamline the haze removal process by using haze state detection to identify which images in the ADAS require dehazing rather than doing this on every image. Then, instead of using a bilateral filter, we used a maximum a posteriori (MAP) with several iterations to refine the fog map and determine the transmission rate and variations in atmospheric light. This eliminates the over-dehazing problem caused by heterogeneous fog in restored images. Finally, we recovered high-quality, haze-free images using an atmospheric scattering model.

The remainder of this paper is organized as follows. Section 2 reviews the extant research and highlights previous contributions which inspired our approach. Section 3 discusses the atmospheric scattering model that is most commonly used for image de-hazing; the linear relationship between brightness and saturation as-applied to identifying haze states is presented, then the fog map is established based on chromatic properties. Section 4 describes the fog map optimization and scene radiance recovery processes. The experimental results are presented and analyzed in Section 5, and conclusions are provided in Section 6.