Wavelet Weighted Distortion Measures for Retinal Images
No Thumbnail Available
This thesis work is an investigation of new distortion measures in the wavelet domain. These measures are expected to quantify the clinical information loss in the processed retinal images. There are three major contributions. First, different subbands are investigated for the presence of retinal image information. Secondly, wavelet weighted distortion measures are defined for each of the retinal features. Third, performance of the new distortion measures is evaluated and compared with that of various state of the art image quality measures. For medical images, the quality measure is expected to emphasize any distortion in the clinically important regions. The traditional squared error measures give equal importance to all parts of the image. To incorporate the importance of distortion in clinically significant regions into the measure, it is required to localize the retinal features such as the blood vessel, optic disc and macula. The spatial frequency localization and multiresolution properties of the discrete wavelet transform (DWT) can make it an efficient way of image representation. The DWT is used to decompose the image into different subbands. The contribution of each subband towards the image information is different. In the present work, two methods are proposed to examine the significance of a subband from the point of view of retinal features. In the first method, different subbands are investigated for the presence of information about the retinal image features. In the second method, the information about blood vessels in different subbands is exploited to propose a new method for the segmentation of blood vessels. After identifying the subbands which contain significant information about retinal features, a wavelet weighted distortion measure is defined. Different weights are assigned to different subband error measures. The loss of clinical information is quantified as a weighted sum of the root of the normalized mean square errors of the coefficients in all the subbands. The distortion measure is computed for each of the features. A global wavelet weighted distortion measure (WWDM) for assessing the clinical quality is then defined by combining the individual distortion measures. The proposed distortion measure WWDM is evaluated quantitatively and qualitatively in predicting the clinical quality of the distorted retinal images. The performance of the WWDM is compared with other image quality measures by applying them on retinal images degraded by different types of artifacts such as JPEG and wavelet compression, blur and noise. The statistical measures used for evaluation are, the Pearson linear correlation coefficient (PLCC), Spearman rank order cor- relation coefficient (SROCC), the outlier ratio (OR) and root mean square error (RMSE) between the actual mean opinion score (MOS) and the predicted MOS values. The statistical behavior is also evaluated in terms of how the measure is discriminating the artifacts when tested on a variety of images using the analysis of variance (ANOVA) method. The WWDM performs better among all the other state of the art image quality measures tested in this work..
Supervisor: S Dandapat and P K Bora
ELECTRONICS AND ELECTRICAL ENGINEERING