engleski

Scene illumination color estimation methods based on convolutional neural networks

Color can be defined as ``the property possessed by an object of producing different sensations on the eye as a result of the way the object reflects or emits light''. Color is a perceptual term that describes the response of the human eye to radiation in the visible range of the electromagnetic spectrum. Due to the varying reflectance properties, different objects in the same scene emit the incident light differently. The human visual system~(HVS) perceives these objects as differently colored. The appearance of an object in terms of color varies depending on the light source in the scene because the perceived color of an object is subject to its reflectance properties and the light source's spectrum. However, HVS is very robust to changes in the observed scene and adapts rapidly. HVS can perceive objects' colors invariant of the present light source ; therefore, a lemon is yellow in, e.g., sunlight and under the light of an incandescent light bulb. This ability is called color constancy. Image sensors in digital cameras, on the other hand, do not have this ability. Therefore, in digital cameras, a pre-processing step is dedicated to achieving invariance of colors to the scene illumination. That step is referred to as computational color constancy. The impact of the illumination color on colors in digital images is usually removed in two steps. First, a method estimates the color of the light source. Second, chromatic adaptation using the estimate renders an illumination-invariant image. The outcome is an image in which white objects indeed appear white. Thus, it is referred to as white balancing. The focus of this thesis is on illumination estimation. By definition, it is an ill-posed problem. Given only image pixels, a vector representing scene illumination has to be estimated ; however, two (or more) individual reflections in the scene can map to the same pixel value. A straightforward approach toward the solution in such tasks is to relax the problem using assumptions. The aim of this research is the analysis of various approaches to illumination estimation. Each approach is implemented in a method using deep learning methodology. Many illumination estimation methods already exist, but often they become inaccurate when more complex scenes occur. On the other side, deep neural networks achieve state-of-the-art results in many computer vision tasks. They were set apart by their exceptional generalization capability. Deep learning methods based on convolutional neural networks accomplish great accuracy in illumination estimation as well. Deep learning usually involves the question of large datasets that are not typical in illumination estimation. A part of the research in this thesis is dedicated to studying existing datasets and establishing a set of desired features a dataset for illumination estimation should have. Finally, a novel dataset conforming to the defined features is presented.

color constancy ; convolutional neural networks ; deep learning ; illumination estimation ; white balancing

nije evidentirano