Bit-depth expansion for noisy contour reduction in natural images

We propose a bit-depth expansion (BDE) method targeting natural images. In the analog part of an imaging system, signal intensity fluctuations occur due to noise (e.g. thermal noise in the image sensor). After that, in the digital part, the intensities are rounded off to limited levels. The latter process, which is quantization, increases the intensity of fluctuation errors caused by stochastic resonance. These errors are viewed as false contour artifacts in the gradation region. Our goal was to obtain the original signal from the quantized noisy signal. We formulated a probabilistic model based on this quantization process, and successfully reconstructed smooth gradations from noisy contours. Subjective evaluation by voting clarified that the output image has higher quality.