Pattern Recognition Letters 2 (1984) 319-327 September 1984
North-Holland
Properties and some fast algorithms of the Haar
transform in image processing and pattern
recognition
P.C. MALI, B.B. CHAUDHURI and D. DUTTA MAJUMDER
Electronics and Communication Sciences Unit, Indian Statistical Institute, Calcutta 700 035, India
Received 4 August 1983
Revised 24 May 1984
Abstract: Properties of the Haar transform in image processing and pattern recognition are investigated. A lower bound of
the performance of the Haar transform relative to that of the Karhunen-Loeve transform for first-order Markov processes
is found. It is proved that the Haar transform is inferior to the Walsh-Hadamard transform for such processes. A unique
condition is presented which, if satisfied by the elements of a matrix, will make the Karhunen-Loeve transform of the matrix
and the Haar transform equivalent. Some fast algorithms are given to realize the diagonal elements of a Haar transformed
matrix.
Key words'." Haar transform, image processing, pattern recognition, Karhunen-Loeve transform, fast algorithm.
1. Introduction
The data compression, feature selection and diagonalisation properties of a sinusoidal family of unitary
transforms have recently been studied and compared with those of Karhunen-Loeve transform (KLT)
(Ahmed et al. (1974), Jain (1979)). A similar study on other fast unitary transforms such as the
Walsh-Hadamard transform (WHT), Haar transform (HT), Slant transform (ST), etc. is also very useful.
Although some practical results have been reported (Pratt et al. (1974)) a detailed analytical study of the
properties appears to be lacking. To this end the present authors reported recently some properties and fast
algorithms on WHT for image models (Mali et al. (1983)).
This paper is an extension of this work for HT. The extension is useful and important because HT may
be fruitfully applied for feature selection, data coding, edge detection and multiplexing problems (Shore
(1973)). Also, HT is locally as well as globally sensitive and is computationally least expensive among the
unitary transforms referred to above.
At first, a relative measure of performance of any unitary transform with that of KLT will be defined. A
lower bound of the relative measure is found for images modeled by a first-order Markov process in Section
2. In Section 3 it is shown that the relative measure for WHT is better than that for HT for any intermediate
case. The image model is represented by a first-order Markov process and both numerical and analytical
results are shown. In Section 4 a unique condition is presented, which, if satisfied by the elements of any
matrix, will make the KLT of the matrix and the HT equivalent. For such a matrix, therefore, the computa-
tionally less efficient KLT is unnecessary. In Section 5 a fast algorithm is proposed to find the diagonal
elements of the two-dimensional HT of any matrix. If the matrix is symmetrical, as is the covariance matrix
0167-8655/84/$3.00 © 1984, Elsevier Science Publishers B.V. (North-Holland) 319