Local Contrast Enhancement with Multiscale Filtering
Kohei Hayashi, Yoshihiro Maeda, Norishige Fukushima · 2023
Multiscale processing is fundamental for image processing. Most multiscale processing splits signals into base and detail signals and then manipulates the detail signal. Local Laplacian filtering (LLF), multiscale processing, has an unusual form: manipulation and then separation. This structure brings out the high performance of the filter. However, LLFs are limited to the form of Laplacian pyramids, which requires downsampling to create image pyramids; it causes misalignments around edges. In this study, we break away from the pyramid and extend the interpretation to multiscale filters that locally enhance the contrast. We propose a local difference of Gaussian (DoG) filter (LDF) for the solution. DoG filtering can remove the downsampling in the filter. Moreover, we further improve the computational speed using the precomputing technique introduced in LLF. Furthermore, to accelerate it, we utilize O(1) Gaussian filtering for large-scale filtering. Experimental results show that the proposed method suppresses overshoot and undershoot around sharp edges caused by the misalignment. Furthermore, the proposed method can keep computational speed even when the convolutional radius is large by O(1) Gaussian filtering.