||Image surveillance is the major means of security monitoring. Image sequences obtained through surveillance cameras are vital sources for tracking criminal incidents and causes of accident, happening mostly at night due to lacking of light and obscurity of vision. The quality of the image plays a pivotal role in providing evidence and uncovering the truth. However, almost all image processing techniques focus on daylight environment, seldom on compensating artifacts rooted from artificial light source at night or light diffusion. The low-lighting environment and color obscurity often invalidate further identification from the surveillance video acquired. |
The processing of images acquired at night cannot follow the paradigm of the daylight image processing. Take image dehazing for example, the removal of haze depends on the derivation of scene depth. Dark Channel Prior (DCP), using dark channel as a prior assumption, is often applied to derive scene depth from a single image. The farthest area, with the highest intensity of light, in an image corresponds to the major source of lighting – daylight, while the area closer with lower degree of light intensity, Therefore, the depth within the scene links with the amount of background light. The above observation does not hold at night. The source of light does not come from sun, rather artificial light source, e.g., street lamp or automobile headlight. The farthest area, often dark-pitch due to lack of any light source, does not have the highest light intensity. To the best of our knowledge, no research has been reported regarding the nighttime image dehazing and enhancement. In light of the demands of higher nighttime image quality, this paper proposes an image dehazing technique, incorporating the light diffusion model, artificial light source, and segmentation of moving objects within the image sequence, to restore the nighttime scene back to the daytime one.
The paper, employing the dehazing and image enhancement to remove the light diffusion in a nighttime image, is composed of daytime background dehazing and nighttime image enhancement. The scene depth is derived by applying DCP to the daytime background image, producing the corresponding depth map. The haze within the scene is removed by the dehazing algorithm to restore the daytime background. The reflectance of objects in the background can be further derived by taking the daylight intensity into consideration. The position and overall intensity of the artificial light sources can be determined through the nighttime background image first. The moving objects are then segmented from the image sequence. The reflectance of moving objects can be evaluated, given the depth map obtained from the daytime image, and position and overall intensity of the artificial light sources from the nighttime counterpart. Once the reflectance of moving objects are determined, the background and moving objects can be fused together given proper daytime lighting.