Enhancing classification accuracy through a cluster-based training approach: A case study of butterfly species classification

In this research work, a method of enhancing the accuracy of classification process is proposed. The method exploits the idea of clustering to choose the right samples to efficiently train the learning models. The work also demonstrates that the segmentation of Region of Interest (ROI) from a given seen image and normalization of the segmented ROI is also very much important to train the learning models efficiently. In order to study the efficiency of the proposed method of training the learning models, a case study of butterfly classification was considered. A butterfly dataset consisting of 3759 images belongs to 15 different species collected through field work and from web sources was created. The images have complex background and significant variation in the foreground (ROI). The traditional machine learning and the popular deep learning models were trained considering both the segmented and non segmented ROI in butterfly images. For traditional machine learning, HOG features were extracted to train the model and SVM classifier was used for classification process. For deep learning, VGG-16, VGG-19, and Inception–V3 models were trained and used for the classification. K-means clustering technique was used to cluster the samples of butterfly images belong to particular class individually. After clustering the samples of butterfly images, one of the image was selected as cluster representative and used for training the learning models. Clusters were formed for different values of K and the K value was chosen depending on the percentage of samples used for training and testing process. Different percentage of training and testing samples were considered to train and test the model. So several experiments considering different percentage of training and testing, such as 90:10, 80:20 etc., were conducted and machine learning and deep learning models were validated using the metrics precession, recall, F1-score and, accuracy. Similarly, several experiments were also conducted for random split of training and testing samples of different ratios. It was observed from the experiments that the cluster based methods of selecting samples for training the learning models was outperformed the random split of training and testing the learning models.