Konfigurasi Optimal Guided Filter dan CNN pada Peningkatan Kualitas Citra yang Memuat DerainNet

Abstract

DerainNet is a Convolutional Neural Network (CNN) based image enhancement method that was designed to remove rainy effects from an image. On DerainNet, an input image was decomposed into base layer image and detail layer image. Base layer image was acquired using fast guided filter as lowpass filter. In this article the authors discuss the effects of using guided filter with multiple configurations of degree of smoothing and neighborhood size as lowpass filter in DerainNet. To see the effects, two assessment methods will be used which is Structure Similarity Index Measurement (SSIM) for synthesized rainy image inputs and Natural Image Quality Evaluator (NIQE) for real world rainy image inputs. The result of DerainNet using the guided filter as lowpass filter will be compared with the result of fast guided filter. Based on the acquired SSIM and NIQE score, guided filter has better results than fast guided filter’s with a SSIM score of 0.919 and NIQE score of 3.829.