Colour image quantisation and coding for optimal perception
Once a digital image is processed in some way and the reconstruction is compared to the original, the final arbiter of reconstruction quality is the human to whom the images are presented. The research presented here is concerned with the development of schemes for the quantisation of colour images and for the encoding of colour images for transmission, with the goal of minimising the perceived image distortion rather than minimising a traditional error signal statistic. In order to quantise colour images with minimum perceived distortion, a colour space is sought in which Euclidean distances correspond linearly to perceived colour difference. The response of the visual system to colour and colour difference is investigated. A new quantisation scheme is developed and implemented to achieve a colour image compression ratio of approximately 6: 1. Three variations on the basic quantiser algorithm are considered and results of applying each variation to three test images are presented. Two-component encoding of colour images for low bit-rate transmission is investigated. A new method of encoding the contents of the image regions following contour extraction is developed. Rather than using parametric surface descriptions, a quad-tree is constructed and a simple measure of perceived image contrast threshold is used to determine the transmitted data. Arithmetic entropy coding is used to discard statistical redundancy in the signal . A colour wash process recreates the colour in each region. Implementation details are presented and several examples are given to illustrate differing contrast thresholds with compression rates of up to 50: 1. An analysis of the textures in certain regions of the test images leads to the development of an algorithm to synthesise the appearance of the textures following extraction of a small block which may be repeated across the region, leading to dramatic compression rates in · some instances.