Image Visibility Enhancement Under The Gradient Domain Processing Framework

Computer vision systems are widely used in many indoor and outdoor scenes including intelligent community, live television broadcast, intelligent transportation, intelligent vehicles, bottom hole survey, night navigation, etc. Visible-light color images are a form of data which is the closest to the human visual perception and are also a form of data input in the large number computer vision system. However, in poor visibility conditions such as at night, haze and fog, the images from the visible-light sensors are seriously degraded with the poor contrast and color fidelity. These degraded images make the computer vision systems unable to work properly, leading to a significant waste of resources. Therefore, it shows important theoretical value and practical significance to effectively improve the image quality of color images collected in poor visibility conditions, which can enhance the robustness and applicability of the computer vision systems.Compared with visible imaging, infrared imaging utilizes each part of difference of thermal radiation of objects to obtain the details of the thermal images. It is not dependent on the ambient light and has powerful ability to penetrate through smoke and hazy. Due to the development of hardware platform, the imaging quality of infrared images has been dramatically improved and the raw data which are acquired from the modern thermal cameras frequently exhibit a high dynamic range (HDR). A high bit depth can potentially accommodate large signal variation, which can obtain an accurate thermal radiometric difference for objects within the monitored scene. However, the majority of display devices only display 256 gray levels which correspond to 8 bits, and they cannot represent all information available in the original data. The poor optical resolution of infrared images has become a bottleneck of further development and application of infrared imaging technology. Thus, establishing visualization technique that effectively enhances the visibility of low-contrast details of HDR infrared images is a very meaningful and challenging work.In order to reproduce the scenery features of the different operating conditions and improve the performance of the computer vision systems including military reconnaissance, safety driving assist system, urban safety monitoring and remote sensing, this research is conducted on three aspects:visibility enhancement of visible images in low-light conditions, visibility enhancement of visible images in foggy and hazy conditions, and visibility enhancement of HDR infrared images. The main research contents are as follows:(1) A comprehensive summary of visibility enhancement methods of the above images is presented and the problems of these visibility enhancement methods are analyzed in this study. Based on the above, a gradient domain processing framework widely used in the processing of image and video is proposed to enhance the visibility of the above images.(2) Traditional contrast enhancement methods cannot simultaneously compress the dynamic range, adjust the brightness and enhance or preserve the details of images on the low-light images processing. By imitating global and local adaptation of human visual system (HVS), a novel contrast enhancement method is proposed to enhance low-light images via a global brightness mapping to local details compensation strategy. Firstly, a nonlinear global brightness mapping model is employed to compress the dynamic range and adjust the overall brightness of the image. Then, the image gradient filed is modified to enhance and restore the local details by combining the luminance masking and suprathreshold contrast perception characteristics of HVS. Finally a new enhanced image is obtained by using a fast Poisson solver on the target gradient field. The experimental results show that our method can compress dynamic range and improve the global brightness effectively. Meanwhile, the proposed method enhance and preserve the details in the dark and light regions, which leads to a great improvement in image visibility, and has no local hazy or halo artifacts which intrinsically belong to other methods such as Retinex.(3) Traditional physics-based single image fog removal methods usually divide by the estimated transmission to obtain the real scene radiance. These methods are prone to introduce "halo" artifacts near the strong edges and cause color over-saturation in the deep depth regions. To overcome these problems, a physical-model-driven defogging method in the gradient domain is proposed in this study. According to the estimated sky light, the white balance is firstly performed to calibrate the fog color to pure white and simplify the atmospheric scattering model. Then the contrast and color of the image is initially reconstructed by subtracting the transmission from the white-balanced image. By introducing some reasonable assumptions, a relational expression between transmission and image gradient is derived. Based on this relational expression, the gradient gain function associating with the transmission is designed to improve the visibility of details. Based on the designed gradient gain function, the small and large scale gradient gains are performed to produce the two images, respectively. The two images can effectively restore the details in the close-range and the distant heavily hazy regions. The final enhancement image is produced from the desired gradient field which is obtained by fusing the gradient of the above two images. Using these synthetic and real foggy images, the performance of the proposed method is subjectively and objectively compared with other methods. The experimental results show that our method can restore the contrast and color of the foggy images effectively and avoid introducing halo artifacts near the strong edges and color over-saturation in the large depth regions. On the other hand, compared with other physics-based single image fog removal methods, the proposed method shows another strength that it can achieve good restoration of contrast and color of the degraded image in rain, snow and underwater conditions.(4) To find the trade-off between providing an accurate perception of the global scene and improving the visibility of details without excessively distorting radiometric infrared information, a novel histogram-statistical-property-driven enhancement method in the gradient domain for HDR infrared images is proposed in this thesis. The proposed method adopts an energy function which includes a data constraint term and a gradient constraint term. In the data constraint term, the classical histogram projection method is used to perform the initial dynamic range compression to obtain the desired pixel values and preserve the global contrast. There are three steps in designing the gradient constraint term. Firstly, taking the mean and standard deviation of histogram projection result as references, the moment matching method is adopted normalize the original HDR infrared image to obtain the normalized image. The normalized image preserves the local structural information of the original infrared image and its dynamic range is relatively consistent with the histogram projection result. Secondly, to achieve parameter adaptability to the different operating scenarios and obtain better visual quality of the resulting image with the proposed method, the percentage of valid gray levels of the original image and the contrast perceptual characteristics of HVS are integrated to design the gradient gain factor function. Thirdly, the desired gradient field is obtained using the designed gradient gain factor function to adjust the magnitudes of the normalized image gradients. Finally, he final output image is obtained by mapping the low-dynamic-range image solved from the proposed energy function linearly to the 8-bit domain. The proposed method is tested by using the infrared images obtained from different operating conditions. Compared with other well-established methods, the proposed method shows a significant performance in terms of dynamic range compression, while enhancing the details and avoiding the common artifacts, such as halo, hazy and saturation. It achieves a high fidelity improvement in infrared image visibility.