Introduction to Transform Coding


One dimensional signals are usually represented in terms of a set of basis functions. The Fourier transform has cosinusoids as its basis functions. We can expand the concept of transforms to 2 dimensions. An image can be represented in terms of a set of arrays called basis images.

The purpose of image coding is to represent the image with a fewer number of bits, while maintaing the visual quality of the image. Image transforms have been widely used to implement image coders. Image transforms are useful for coding, since images generally have a compact representation in the transform domain.

An image transform takes N*M pixel values in an image and transforms them into N*M transform values. The original image has its energy spread throughout the image. This means that the large pixel values are likely to be anywhere in the image. A "good" transform has a high degree of energy compaction. The large transform values are localized. The histogram of pixel values illustrates the fact that most pixels have very low values and a few pixels have large values. A large number of bits are allocated to these few transform coefficients. Very few bits are allocated to the small transform values. Thus most of the information in the image can be represented by a smaller number of total bits.



The compressed image can be obtained by applying the inverse transform to the transform values. The optimal transform is the one that minimizes the distortion between the original and compressed images. This transform is the Karhunen-Loeve transform. However the Karhunen-Loeve transform requires knowledge of the second order statistics of the image and is very difficult to implement. Since the second order statistics of the image are difficult to obtain, approximations to the Karhunen-Loeve transform are preferred in practice. One of these is the Discrete Cosine Transform (DCT). The DCT is a fast transform with complexity of O(NlogN). It gives real values which correspond to the frequency content of the signal. It works extremely well with highly correlated data. One versionof the DCT is implemented in the popular JPEG compression scheme. The wavelet transform is another excellent transform for image coding. Its results are closer to that of the K-L transform than the Discrete Cosine Transform.



Back..