Perceptual Image Enhancement And Its Applications To Display Enhancement And Visual Surveillance

Nowadays,images and videos are widely used in various fields of social life such as intel-ligent transportation,medical diagnosis,target detection,and human-computer interaction In display intelligence and visual surveillance,the visual quality of images is always cor-rupted by the bad viewing(e.g.strong ambient light)and imaging conditions(e.g.low light or haze).Thus,it is necessary to apply image enhancement to degraded images,which improves the image quality.However,conventional image processing cannot adaptively control the enhancement degree according to the viewing and imaging conditions.And it al-ways causes over enhancement,noise amplification and detail reduction in resulted images Thus,we want to develop perceptual image processing which models and utilizes human visual characteristics.These characteristics are applied to adaptively improve image quali-ty and prevent severely over enhancement.In this thesis,we investigate the image quality degradation problems in display enhancement and visual surveillance,and develops the per-ceptual image enhancement based on image characteristics and HVS models.The thesis is organized as follows:(1)Ambient light adaptive display enhancement:Image quality in displays is consider-ably influenced by ambient light.In the daylight condition,images on displays are darkly perceived by the HVS due to the limited dynamic range of display,which causes loss of luminance,contrast and details.Therefore,we propose the ambient light adaptive display enhancement to improve the readability of display images(a)Inspired by the luminance and detail distortions under ambient light,we develop the optimized luminance and detail enhancement of images in displays under ambient light This method improves the readability of the displayed image by simultaneously enhancing luminance and detail of the image using the lightness adaptation model and contrast sensitive model(b)However,single luminance enhancement by transfer function(i.e.gamma function)always leads to dynamic range compression to high intensity range.The enhanced images look foggy without clear structure under ambient light.To achieve the enhanced displayed with high luminance and contrast,we propose adaptive tone mapping using constrained optimization under ambient light.The proposed method adaptively control the enhanced degree which depends on the lightness adaptation model and ambient light,and achieve the enhanced images with high contrast and minimum distortions(2)Optimized contrast-tone mapping of displayed images for backlight dimming:Al-though backlight scaling is an effective tool for saving power,it inevitably causes brightness reduction,which degrades the displayed image quality.We propose power constrained con-trast enhancement based on brightness compensated contrast-tone mapping operation(BC-CTMO).We adopt brightness compensation to restore the original perceptual luminance,and combine tone mapping with contrast enhancement to solve the un-displayable problem caused by brightness compensation,which improves image quality(3)Perceptual low light image enhancement with two-step noise suppression:In the previous three works,the proposed method focuses on the display enhancement and the o-riginal images have good image quality.However,low light images suffer from low contrast and serious noise.Thus,we propose the low light image enhancement based on perceptual two step noise reduction.In the first step,we develop noise aware contrast enhancement based on noise characteristics and image locality to prevent noise amplification after con-trast enhancement.Then,we have perceptually reduced noise in images while preserving details using a JND model which represents noise visibility in contrast enhancement(4)Multi-spectral RGB and NIR fusion using multiscale wavelet analysis:Multi-spectral fusion of RGB and near-infrared(NIR)images is a novel approach which achieves more attention because non visible light imaging including NIR provides fine details and clear structure especially in low light conditions.Thus,we develop the multi-spectral fusion and denoising(MFD)framework of RGB and NIR images using multi-scale wavelet analysis To deal with discrepancy between RGB and NIR images,we conducted the wavelet scale map estimation.We have selected high and visible contrast pixels from NIR image and transferred them to RGB image.Then,we have adjusted NIR image based on the scale map to fuse it with RGB image,and have used the adjusted NIR image as the guidance map to reduce noise in the fusion result.The proposed method achieves high quality images with fusion of NIR and RGB images in the results.