Object Recognition and Segmentation of Wounds

Abstract

Object recognition and segmentation of objects is a complex task. Our goal is to develop an algorithm that can recognize and segment wound objects in images. We attempt to solve the object recognition and segmentation problem by using a hypothesis optimization framework. This method optimizes the object segmentation by assigning objective function values to the object segmentation hypotheses. The optimization algorithm is a genetic algorithm. The objective function relies on textural and shape properties, and the textural properties relies on classification of superpixel-segments and superpixel-edges within wound images. Superpixel-segments and superpixel-edges within the same image are dependent samples. We use combined hyperparameter and feature selection methods to train classification models, and we evaluate the impact of dependent samples on these methods. To our knowledge, no study has evaluated model-selection methods when the data contains known groups of dependent samples. Our results confirm that dependent samples results in biased error estimates. Biased error estimates can cause suboptimal feature and hyperparameter selections, and therefore reduce the classification performance. Finally, we obtain promising results by using hypothesis optimization to solve object recognition and segmentation of wounds. These results are important because of the flexible nature of hypothesis optimization; they demonstrate that hypothesis optimization is a strong candidate for general-purpose machine-learnable object recognition and segmentation.