Development of a pattern recognition framework focusing on feature extraction and selection on images

Nowadays, the number of digital images is increasing gradually with the use of technological possibilities and devices. Computer aided classification of image types is widely applied in many applications such as medicine, security, automation. The feature extraction and selection stages have great importance in terms of increasing the classification performance as the substages of the pattern recognition process. In the thesis, a pattern recognition framework including both of feature extraction and selection stages on images is developed. In the first submodel of the framework, only manual feature extraction methods are used. In this model, making up a large set of methods for extracting 194 features, 4 of which are novel ones derived from images divided into two sections according to the center point, is one of the contributions to the literature. Genetic algorithms are utilized in the feature selection stage. In the second submodel, which is developed as another contribution to the literature, the system is extended by adding the feature set extracted by using deep learning to the first submodel, thus recognition performance is improved. In the scope of the thesis, a framework was created for extracting total of 4290 features (194 handcrafted and 4096 deep learning features), Flavia and Caltech-101 datasets were used in the experimental studies. It is shown that the best classification performance for two datasets is provided with the ECOC-SVM model and it is competetive compared to the existing state-of-the-art results in the literature.