STUDY OF IMAGE QUALITY IMPROVEMENT TECHNIQUES

Abstract

Image quality improvement has always been a topic of interest for the researchers. The ultimate objective of image quality improvement in a broad sense is to improve a degraded image that can express all the information of the scene. The need for the image quality improvement is necessitated because of the limited hardware capabilities of image capturing devices, the uneven lighting conditions, and noisy environment. In this scenario, post-processing is needed to improve the quality of the acquired image. This thesis explores new algorithms for image quality improvement. This thesis addresses several open questions: a) How do the repetitive geometric structures such as ridges and valleys in fingerprints can be exploited for better sparse representation of fingerprint images? b) How can denoising and superresolution algorithms be incorporated in traditional enhancement methods to help in reliable and accurate fingerprint matching? c) How can histogram equalization be effective in contrast enhancement of low exposure images? In an attempt to answer these questions, this thesis explores three aspects of image quality improvement i.e. denoising, super-resolution and contrast enhancement. Being motivated from recent advancements in sparse coding based image processing applications, a novel ridge orientation based clustered sparse dictionary is proposed for exploiting self-similarity in fingerprint images which often contains many repetitive geometric structures such as ridges and valleys. Instead of having a single dictionary, the proposed dictionary-learning method clusters the training patches into smooth, non iii dominant orientation and dominant orientation groups. The use of sub-dictionaries based on dominant orientation best describes the underlying image data. This also improves the effectiveness of sparse modeling of information in a fingerprint image in the form of local ridge patterns and thus further improves denoising and super-resolution performance. Two novel fingerprint image denoising and super-resolution algorithms based on ridge orientation based clustered sparse dictionary are proposed. The fingerprint denoising approach undergoes three steps i.e. Ridge orientation based patch clustering, Sub-dictionary learning and sparse coefficient calculation. While reconstructing the denoised image in the final step, the minimum residual error criterion is used for choosing sub-dictionary that best suits for a particular patch. The fingerprint super-resolution algorithm involves learning of coupled sub dictionaries each for low and high-resolution training patch groups that are clustered based on dominant orientation. In the final step of superresolution, the iterative back projection is applied to eliminate the discrepancy in the estimate due to noise or inaccuracy in the sparse representation. To evaluate the performance of proposed fingerprint denoising and super resolution methods, a validation methodology consisting of three experiments i.e. comparison based on image quality measures, visual quality, and fingerprint matching is devised. The performance comparison results show that the proposed methods achieve better results in comparison with other methods and will help in improving the performance of fingerprint-identification systems. iv This research also addresses the problem of low exposure imaging through contrast enhancement using histogram equalization. A novel Exposure based Sub-Image Histogram Equalization (ESIHE) method for contrast enhancement is proposed which partitions the image into sub-images i.e. under-exposed and over-exposed. The individual histograms of sub-images are equalized independently. The two recursive variants of ESIHE are also proposed that proves to be very effective for improvement in the quality of images acquired in low light conditions such as underwater sequences or night vision. The first method is Recursive Exposure based sub-image histogram equalization (R ESIHE) that recursively performs ESIHE method till the exposure residue among successive iteration is less than a predefined threshold. The second method is named as Recursively separated Exposure based sub-image histogram equalization (RS-ESIHE) that partitions the image histogram recursively. Each sub-histogram is further partitioned based on their respective exposure thresholds, and finally all the sub-histograms are equalized individually. The experimental results show that the proposed methods efficiently handled the low exposure image enhancement problem that was not addressed by earlier HE based methods. Another variant of histogram equalization i.e. Median-Mean based sub-image clipped histogram equalization (MMSICHE) method is developed to addresses the problem of preservation of mean brightness, entropy and control on the enhancement rate simultaneously. The algorithm consists of three steps, namely median and mean calculation, Histogram Clipping and Histogram Subdivision & Equalization. The important factors for accomplishing the objective by MMSICHE method can be summarized as (i) Bisecting the image based on the median value plays the role in v maximizing entropy and natural enhancement. (ii) Sub-dividing the image into four sub images based on mean intensity plays the role for brightness preservation. (iii) Histogram clipping approach provides the feature of control on over enhancement. The simulation results show that MMSICHE method avoids excessive enhancement and produces images with a natural appearance.