Super resolution method for the enhancement of low quality images

Abstract

Super resolution refers to the process of obtaining a high resolution image from a single or multiple low quality image/s. Super resolution has become a very important area of research in the Engineering society. This is due to the fact that there is always such a high demand for a great deal of detail in images from various different industries. Such industries include space exploration, medical imaging, forensic imaging and various other image/video applications, such as the initial phases for a face detection and recognition algorithm. The better the quality of the data supplied to such industries, the better would their end result be. It is therefore imperative that these industries obtain very accurate images so as not to hinder results. This dissertation presents a super resolution method of a single low resolution image, while maintaining a high level of information in details, features and edges. The algorithm presented is based in the wavelet domain. This is because in the wavelet domain, edges and details are preserved better when compared to spatial interpolation techniques. Most spatial interpolation techniques have a tendency of blurring the image since they are not able to differentiate between the different patterns of the image. On the contrary, by exploiting the way coefficients are stored in the wavelet domain, the system is able to differentiate between the different types of patterns in the image, and train these components accordingly. These components are mainly the coefficient approximation (LL), the horizontal detail (LH), the vertical detail (HL) and the diagonal detail (HH). The wavelet transform was done using the Haar filter and the CDF 917 wavelet transform to obtain the HL, HH and LH components. Each of these components were divided in 4x4 blocks, which were then categorised into two regions. The two regions represent either a smooth region or else a block which represented an edge. This classification depends on the mean of the block. Six neural networks were then trained, two for each of the detail co-efficient. The CDF 917 wavelet transform gave a very good subjective result but a poor PSNR. On the other hand the Haar wavelet transform returned a good result both subjectively and quantitatively.