Gradient domain high dynamic range compression
Proceedings of the 29th annual conference on Computer graphics and interactive techniques
Fast bilateral filtering for the display of high-dynamic-range images
Proceedings of the 29th annual conference on Computer graphics and interactive techniques
Distinctive Image Features from Scale-Invariant Keypoints
International Journal of Computer Vision
Compressing and companding high dynamic range images with subband architectures
ACM SIGGRAPH 2005 Papers
| Hi-index | 0.00 |
A novel approach to a virtual high dynamic range (HDR) imaging for the object recognition robust to large radiance variations is presented. Unlike the conventional approaches to HDR imagery and tone mapping [10] that seeks for the enhanced displays optimal for human perception, the proposed approach aims at the robustness in recognition under the presence of poor and/or saturated radiance in imaging. The method first transforms a low dynamic range (LDR) image under the existence of poor and/or saturated recordings into a virtual HDR image with no poor and/or saturated ones. The key to the proposed HDR transformation is extending the shutter speed and intensity relationship of individual pixels beyond saturation, in the real color space, based on an LDR image captured. The method then applies a histogram optimization to the HDR image obtained, in the intensity space, in such a way as to fit it into an LDR image suitable for processing. A mathematical formula is presented as a general form of histogram transformation. The proposed histogram optimization then determines the parameters of the formula in such a way that the resulting image can deliver the maximium information for recognition. An extensive experimentation is conducted to evaluate the impact of the proposed virtual imaging on the robustness in recognition against large radiance variations, in particular, using the Scale Invariant Feature Transformation (SIFT) [13] as a testbed for recognition. To provide a quantitative measure, a performance comparison among the following three imaging modalities is conducted: 1) the imaging with the camera auto exposure only, 2) the imaging by the combination of the camera auto exposure with the conventional histogram equalization, and 3) the proposed virtual imaging with the HDR histogram optimization. The result demonstrates that the proposed virtual HDR imaging is superior in performance to the counterparts, providing the capability of recognizing an object not only under its highly poor radiance but also under its highly saturated radiance imaging with the camera auto exposure only.