A novel machine learning approach in Image Pattern Recognition under invariance constraints

Two of the main challenges of image recognition in radar, acoustic or T-ray imaging regard the view-point variation of the pattern and the feature extraction techniques that must retrieve the most discriminative information about different classes. In this paper, we focus on feature extraction and image classification techniques by using a Rotation Invariant Wavelet Packet Decomposition and a novel entropy-based feature extraction technique to characterize an image. The entropy-based characterization described in the paper offers an extended analysis compared to usual approaches such as the energy of the wavelet sub bands. The computed features will be further used to train a Graph Neural Network adapte d to a quad-tree decomposition which has the powerful advantage of considering the structural information of the rotation-invariant decomposition. We successfully classified the images with an accuracy of 99.3%. The results are compared to other classic feature extraction techniques such as k-NN, SVM, DCT and WPD, proving the increased capability of our method.