No-reference and no-training based image quality assessment and enhancement

Abstract

There are chances of images getting distorted while passing through several op- erational stages such as image acquisition, compression, transmission, processing and reconstruction. Use of distorted images in any application degrades its performance; and therefore, identi cation and quanti cation of the distortions present in the image become very necessary for controlling and enhancing the image. This has motivated us to propose image quality assessment and enhancement techniques for various kinds of distortions. The proposed techniques in the thesis function in two steps. In the rst step, quality of distorted images is estimated whereas in the second step, enhancement is performed on distorted images based on the obtained quality in the rst step. The proposed work considers three important distortions, viz: contrast, illumination and noise for quality estimation and enhancement of an image. Along with these distor- tions, the thesis also proposes a quality assessment technique for deblocked images. All the techniques proposed in the thesis are based on no-training and do not require any reference image for quality assessment or enhancement.For contrast based quality assessment, the proposed technique in the thesis utilizes local band-limited contrast at multiple spatial frequencies. For enhancement of the contrast, the thesis presents a technique based on arti cial bee colony (ABC) opti- mization. This technique proposes the inclusion of direction constraints in ABC to make arti cial bees to move in right direction to obtain better solution and to reduce computational time. It further proposes the use of contrast based quality estimation as an objective function in ABC.The technique proposed for illumination based quality assessment utilizes the con- cept of Retinex theory to compute quality of an image. The technique computes quality by rst identifying dark and bright regions present in an image and then cod- ing the pixels of these regions based on their neighbors. For improving the quality of the image having non-uniform illumination, the technique proposed in the thesis goes through following steps. In the rst step, it proposes a bad illumination pass lter to locate poorly illuminated areas in the image and then generates an image with goodillumination by removing poorly illuminated areas. In the second step, it proposes a quality aware logarithm based enhancement technique to insert proper amount of brightness in the image in non-uniformly illuminated areas. There are two major con- tributions of this work. First contribution lies in proposing a bad illumination pass lter to locate poorly illuminated areas in the image and subsequently creating an enhanced image by highlighting details as well as good illumination areas. Second contribution lies in proposing an adaptive logarithm transformation using estimated quality of an image to add the required amount of brightness in poorly illuminated areas in the image. Quality of a noisy image is estimated using singular value decomposition (SVD). The proposed technique is based on two characteristics of the human eye (retina), viz: presence of center-surround receptive eld and visualization utilizing di erent spatial frequency channels. In the technique proposed in the thesis, modeling of center- surround receptive eld is carried out using di erence of Gaussians (DoGs) whereas to mimic multiple frequencies in center-surround receptive eld, multiple DoG images of di erent standard deviations obtained for di erent frequencies are computed. Further, SVD based features are obtained from the generated DoG images to estimate the image quality. After quality estimation, quality aware Wiener lter is proposed to reduce noise in the image. JPEG is considered to be one of the most commonly used compression standard whose resultingimages are found to be subjected to blocking artifacts at low bit rates. There exist a few deblocking algorithms in the literature to reduce the blocking ar- tifacts in compressed images. The thesis proposes a no-reference based technique for assessment of quality of deblocked images which can be used to analyze the perfor- mance of a deblocking algorithm.