A gene expression programming approach for evolving multi-class image classifiers

This paper presents a methodology to perform multi-class image classification using Gene Expression Programming(GEP) in both balanced and unbalanced datasets. Descriptors are extracted from images and then their dimensionality are reduced by applying Principal Component Analysis. The aspects extracted from images are texture, color and shape that are, later, concatenated in a feature vector. Finally, GEP is used to evolve trees capable of performing as classifiers using the features as terminals. The quality of the solution evolved is evaluated by the introduced Cross-Entropy-Loss-based fitness function and compared with standard fitness function (both accuracy and product of sensibility and specificity). A novel GEP function linker Softmax-based is introduced. GEP performance is compared with the obtained by classifiers with tree structure, as C4.5 and Random Forest algorithms. Results show that GEP is capable of evolving classifiers able to achieve satisfactory results for image multi-class classification.