Triplet-CSSVM: Integrating Triplet-Sampling CNN and Cost-Sensitive Classification for Imbalanced Image Detection

In real-world applications, image classes are often imbalanced, which result in detection performance decline and quite different misclassification costs. In order to deal with these issues, cost-sensitive learning based on manually designed features has been studied for many years. With the rapid development of Deep Learning, more comprehensive methods, such as CNN and RNN, have proven their strength on feature extraction and classification. In this paper, we develop triplet-sampling CNN to automatically obtain a great many in-depth features from images. Cost-sensitive SVM (CSSVM) is applied to deal with the classification performance degradation caused by imbalanced image dataset. Furthermore, two techniques are integrated as Triplet-CSSVM for classifying images accurately even over imbalanced image set. This approach can overcome the disadvantages of the conventional features extraction and improve the overall classification performance comparing with several other related schemes.