|
Pixel domain interpolation is the typical method of resolution enhancement of digital images. After enlarging image with interpolation operation, sharp edges usually become aliased. By detecting the edges and compensating, profiles of the interpolated edges can be much smoother. However, the quality of edge detection and compensation method strongly affects quality of the enhancement. Besides, the computation is much complicated. A new approach of resolution enhancement by using the property of Multi-Resolution Analysis through Discrete Wavelet Transform is proposed. By estimating the higher resolution wavelet coefficients, the resolution can be increased by two times through the synthesis operation of Discrete Wavelet Transform. The estimation is done by using neural networks combined with a simple edge classification method to improve the estimation accuracy. Both 1-D and 2-D cases are shown in this thesis. The experiments show that the enlarged images are clear and sharp. Some details are preserved in the processing. However, parts of the edges are sharp unduly and spurious noise is generated. Further more, the training of the neural networks is very slow due to the huge size of training samples. It is desired to overcome these problems in the future.
|