Airborne pollen texture discrimination using wavelet transforms in combination with cooccurrence matrices

In the recent years image processing and pattern recognition techniques have been applied to develop intelligent systems for both of fresh and fossil pollen grains discrimination. In this paper we aim at the texture identification of pollen surface images. A method of texture description using wavelet transforms in combination with cooccurrence matrices is presented, and a neural network is used to classify the extracted image features. In this combined method, through wavelet decomposition and reconstruction, an approximation image and a new details image are generated for the input image. The surface texture of pollen grains is characterised by using a rotational invariant feature set, which is formed from the joint distribution of the grey level and the details information. In order to form effective feature vectors, the moment invariants also were employed to describe the surface shape of pollen grains. Both the back-propagation (BP) and the learning vector quantisation (LVQ) networks were used for classification of the resulting feature vectors. In experiments with sixteen types of airborne pollen grains, more than 91% pollen images are correctly classified using both the methods.