A neural network based predictor of filtering efficiency for image enhancement

Abstract

Image filtering is widely used in remote sensing applications to improve obj ect visibility or for other purposes.However, filtering does not always occur efficient enough and serving image enhancement purposes well. Thus, it is reasonable to have a simple but rather accurate predictor of filtering efficiency. Such a predictor can be based on statistics of DCT coefficients in image blocks. For improved prediction, we propose to apply several local statistics aggregated by a trained neural network. This way allows providing high accuracy prediction of image enhancement not only in terms of standard quality criteria but also in terms of metrics of image visual quality.

Efficiency of texture image enhancement by DCT-based filtering

“…Then, it is desirable to have estimates (prediction) of denoising efficiency for existing filters. The main idea of such a prediction is the following [27][28][29]. Suppose that it is possible to quickly estimate one or several statistical parameters of a given image that are quite strictly connected with a parameter (or parameters) that characterize denoising efficiency.…”

“…Following an analysis in [27][28][29], it is better to use 2 P than 7 . 2 P (prediction is more accurate).…”

“…It is not necessary to use in estimation all possible positions of the image blocks. Usually it is enough to have 500 non-overlapping or partly overlapping blocks to obtain an estimate, which is accurate enough for our purpose [29]. The reasons for this will become clear later.…”

“…First steps forward in attempts to answer this question have been made in the papers [18,[27][28][29]. In [18,27] the authors have shown that there is a connection between simple statistics of DCT coefficients in 8x8 blocks and filtering efficiency.…”

“…It has been demonstrated [25,28] that other parameters describing denoising efficiency such as improvement of PSNR (IPSNR) and improvement of PSNR-HVS-M (IPSNR-HVS-M) (where PSNR-HVS-M is visual quality metric [30]) can be predicted as well. Moreover, very accurate prediction can be achieved if a trained neural network (NN) is applied for approximating the dependence between statistical parameters of a noisy image and a parameter characterizing denoising efficiency [29]. Although textural images were present in the training set, thorough analysis of prediction accuracy just for texture images has not yet been performed.…”