Boosting Performance of SVM in Koi Classification Using Direct Methods-Based Optimization

Abstract

Many koi fish enthusiasts keep or buy them just for their attractive colors without knowing what type of koi fish they are. The manual classification of koi fish species is still frequently incorrect. As a result, it is critical to apply a machine learning technique to identify various koi fish species. This research implemented a computer vision algorithm to classify koi fish species using the Support Vector Machine (SVM) as the classifier. However, the maximum accuracy SVM can achieve in our koi fish classification system is 79%. To achieve better performance, the SVM was optimized by applying various optimization methods from the Direct Method group, i.e., the Generalized Pattern Search (GPS), the Powell method, and the Nelder-Mead method. Three optimization methods from the Direct Method group have successfully improved the performance of SVM in this task. Experimental results demonstrated that using the Generalized Pattern Search (GPS) in our classification system can increase the accuracy to 98%. Also, implementing the Powell and the Nelder-Mead method can make the koi classification system obtain a better accuracy of 99%. These results indicate that the proposed approach is a viable solution to overcome the limitations of the SVM algorithm.