| Summary: | 碩士 === 中華大學 === 資訊工程學系 === 107 === In the past few years, deep learning has become prevalent in computer vision. The detected object affected by the chromatic illuminant will cause identification errors, so it is important to perform color recovery. In previous studies, there were various methods to recover image color. The most useful one is the Maximum-Spectral-Value (MSV) color recovery method. However, MSV method has two limitations. First, the hue and saturation of the illuminant must kept the same over the scene. Second, no backlighting exists in the scene when taking the picture.
Image color recovery is based on the illumination model. The light source is reflected by the surface of the object and perceived by human cone. Mathematically, the reflection of the light spectrum becomes the color value in the image through color matching function. These reflected spectrum can be approximated by using the weights and basis functions of light source and reflectance via principal component analysis. Because MSV method will be affected by the reflection spectrum of high brightness, image color under the non-uniform chromatic illumination and backlighting cannot be recovered.
In this paper, we propose an improved MSV method called Multilevel-Maximum-Spectral-Value method (MMSV) to overcome these problems. MMSV method first divides the image into multilevel regions based on brightness, and then calculates the maximum reflected light spectrum for each region respectively. Therefore, MMSV method is not affected by the other higher brightness region and backlight spectrum can be calculated correctly on each region.
Four typical images under four kinds of backlight source (red, green, blue and white light) and three types of chromatic illuminant (red, green, and blue) are simulated, so there are totally 48 sets of biased color images tested in the experiments. From the experimental results, it is proved that MMSV method can recover images with backlighting and chromatic illuminant. In addition, we have actually taken photographs of common stairwells and cavalry images for recovering of backlight. From the experiments, the proposed MMSV method can recover color effectively and be used in real application.
In some cases, MMSV method may not work well due to huge variations of light source. We also discuss and analyze the effect of intensity of backlight and distribution of chromatic illuminant in the experiments to clarify the limitation of MMSV method. In the near future, we will continue improving the method to reduce the limitation and recover color in more actual conditions, and then apply them flexibly in daily life.
|
|---|
