Guided Filter Bank

Hui Yin, Yuanhao Gong, Guoping Qiu

Abstract: Guided filter can perform edge-preserving smoothing by utilizing the structures of a guidance image. However, it is difficult to obtain two images with different contents from the same scenario. Therefore, we focus on the case that the input image and the guidance image are identical. In this case, the direction of the gradient of the output image is the same as the guidance image. Based on this discovery, we change the regularization term of guided filter and develop a more general model which can generate a bank of guided filters. To take examples, we pick up three filters from this bank, where \(L_1\) guided filter and \(L_{0.5}\) guided filter are newly proposed filters. Mathematical and experimental analysis are performed to demonstrate that the new filters have totally different properties from the guided filter. \(L_1\) guided filter is very suitable for edge-preserving and texture-removing tasks, while \(L_{0.5}\) guided filter can do enhancement automatically. We applied them to a variety of image processing applications, including image smoothing, image denoising, edge detection, image detail enhancement and X-ray image enhancement. The experimental results reveals the effectiveness of the newly proposed filters and their state of the art performance. We also believe that more interesting filters can be developed from our model.