Object Recognition through Turbulence with a Modified Plenoptic Camera

摘要

Atmospheric turbulence adds accumulated distortion to images obtained by cameras and surveillance systems. When the turbulence grows stronger or when the object is further away from the observer, increasing the recording device resolution helps little to improve the quality of the image. Many sophisticated methods to correct the distorted images have been invented, such as using a known feature on or near the target object to perform a deconvolution process, or use of adaptive optics. However, most of the methods depend heavily on the object's location, and optical ray propagation through the turbulence is not directly considered. Alternatively, selecting a lucky image over many frames provides a feasible solution, but at the cost of time. In our work, we propose an innovative approach to improving image quality through turbulence by making use of a modified plenoptic camera. This type of camera adds a micro-lens array to a traditional high-resolution camera to form a semi-camera array that records duplicate copies of the object as well as "superimposed" turbulence at slightly different angles. By performing several steps of image reconstruction, turbulence effects will be suppressed to reveal more details of the object independently (without finding references near the object). Meanwhile, the redundant information obtained by the plenoptic camera raises the possibility of performing lucky image algorithmic analysis with fewer frames, which is more efficient. In our work, the details of our modified plenoptic cameras and image processing algorithms will be introduced. The proposed method can be applied to coherently illuminated object as well as incoherently illuminated objects. Our result shows that the turbulence effect can be effectively suppressed by the plenoptic camera in the hardware layer and a reconstructed "lucky image" can help the viewer identify the object even when a "lucky image" by ordinary cameras is not achievable.