Leaf classification from binary image via artificial intelligence

• A TSO and affine invariant systems of 2D binary image descriptors was developed.
• Properties of invariant descriptors experimentally verified with digital leaf images.
• Ability of invariant systems to retrieve the TSO or affine transformed images proved.
• 1-NN and Support Vector Machine used for leaf classification.
• Kohonen Self-Organising Maps used for leaf categorisation.

The invariant recognition of 2D binary images is the main subject of the paper. Two methods for invariant pattern recognition based on 2D Fourier power spectrum with guaranteed translation invariance are proposed. First method introduce the features invariant to translation, scaling, rotation and mirroring (TSO invariance). The second method introduces the features invariant to general affine transform (A invariance). The methods are used to obtain TSO/A invariant spectra except of the rotation effect which are analysed on circular paths with fixed radii. Harmonic analysis of power fluctuations around paths generates Fourier coefficients and their square absolute values are used as TSO/A invariant descriptors. The proposed methods were tested on two large sets of 2D digital images of tree leaves. After TSO/A invariant processing of thresholded digital images, kernel Support Vector Machine or self-organizing neural network were used for leaf categorisation.