In order to develop vision systems that present under all weather conditions, it is essential to model the visual effects of the various weather conditions and develop algorithm to remove them. Weather conditions vary widely in their physical properties and produce visual effects in images. Based on the types and size of the particles and their concentration in space [11] as illustrated in Table 1.1, weather conditions can be typed as steady and dynamic. In the case of the steady weather, individual droplets are too small (1-10 μm) to be visible to a camera, and the intensity at a pixel is caused by the aggregate effect of large number of droplets within the pixel’s solid angle such as fog, mist, and haze. The haze image model proposed in [1], [2], [11], [12] and [13] is widely applied in computer vision and can be adequately described the effects of steady weather, as
where (1.1) is equal to (1.2), I is the observed intensity of the haze image, x is the pixel’s index, J (i.e., J L( )x with ρ being the reflectance of an object in the image) is the scene radiance of the haze-free image which is desired to be obtained by dehazing techniques, t (i.e., t e d x( ) with β being the atmospheric attenuation coefficient) is the medium transmission expressed part of light that is not scattered and reaches the observer, and A (i.e., A L ) is the global atmospheric light. The first term is the direct attenuation D x (a( ) D xa( )L( )x ed x( ) J t (x)), and the second term is
the airlight A x (r( ) A xr( )L(1ed x( )) A(1- (t x))). Note that the haze removal technique is to estimate the global atmospheric light (A) and the medium transmission (t) from the haze image model (I), and the scene radiance (J) can be estimated by these estimated components. Fig. 1.1 shows the pictorial description of the direct attenuation
a( )
D x and the airlight ( )A x . In (1.1), it geometrically means that in RGB color space, r the vector I is a linear combination pf the two vectors J and A. (see Fig. 1.2). In Fig. 1.3, it depicts the relation between ( ), ( ), and ( ) I x J x t x in (1.1) to corresponding images.
Condition Particle Type Radius
m Concentration
cm3
Table 1.1 Weather condition and associated particle types.
Fig. 1.2 Haze image formation model. The color vector I is a linear combination of J and A in the RGB space.
( ) ( ) (1- ( )) I x t x J t x A
Fig. 1.3 Relation between the functions and its corresponding images in (1.1).
In the other case of the dynamic weather, individual particles are visible in the image (0.1−10mm) and the aggregate scattering models used for steady conditions are not applicable in here such as rain, snow, and hail. But, when raindrops and snow are very far from the camera, their visual effects are very weak and appear as fog. Therefore, what we focus on are rain and snow that are close to the camera and fall on a car windshield. The analysis of dynamic weather conditions requires the development of stochastic models that capture the spatial and temporal effects of a large number of
particles moving at high speeds as in rain and complex trajectories as in snow. Fig. 1.4 compares the differences between the steady and dynamic weather. The change in intensity produced by a falling raindrop as a function of the drop’s distance z from the camera. The change in intensity I does not depend on z for drops that are close to the camera (z z m). While for raindrops far from the camera (z z m), I decreases as 1/ zand for distances greater than Rzm(where R is a constant), I is too small to be detected by the camera, only produce aggregate scattering effects. Therefore, the visual effects of rain are only due to raindrops that lie close to the camera (z Rz m) which we refer to as the rain visible region. The value R depends on the brightness if the scene and camera sensitivity. However, raindrop is considered to be a water droplet on a car windshield which can distort the view especially during driving. Depending on the speed, raindrop flies upward as the speed increasing and flows downward as the speed is slowing down. Nevertheless, water droplet on dirty windshield is not in a spherical form is why the view becomes blurred, and the circumstance is known as unfocused raindrops. Focused raindrops can be seen in a spherical form on windshield as shown in Fig 1.5(a) and unfocused raindrops which cause blurring view as shown in Fig 1.5(b) are vice versa. Both types of raindrop lead to challenging tasks in terms of detecting and removing the raindrop.
m mm
I
z
z
mRz
mFig. 1.4 Comparison of the differences between the steady and dynamic weather.
(a) (b)
Fig. 1.5 Raindrops fall on a car windshield. (a) Focused raindrop. (b) Unfocused raindrop.
Different from common images, underwater images suffer from poor visibility due to the attenuation of the propagated light. The light is attenuated exponentially with the distance and depth mainly due to absorption and scattering effects. The absorption substantially reduces the light energy while the scattering causes changes in the light direction. The random attenuation of the light is the main cause of the foggy appearance while the fraction of the light scattered back from the medium along the sight considerably degrades the scene contrast. By the way, dehazing techniques [14] have been related with the underwater restoration problem.
The remaining of this thesis is organized as follows. In Chapter 2, we introduce three existing typical single image dehazing methods. In Chapter 3, we propose a robust and effective dehazing method based on human vision to improve the dehazing quality during daytime and nighttime. Four existing typical rain and snow removal methods are also introduced in Chapter 4. For rain and snow removal investigations, we design a simple but effective rain or snow removal method by divide the rain or snow removal scheme into two parts, the first part is detection of rain or snow and the second part is inpainting in Chapter 5. In Chapter 6, we introduce three existing typical underwater enhancement methods. In Chapter 7, we propose an effective enhancing method to improve the underwater image quality. Chapter 8 concludes this thesis and discusses the future work.