Handling Highly Imbalanced Dataset on Image Segmentation Using Ternary Classifier

Abstract

A traditional problem in computer vision, image segmentation refers to segmenting the area of an object existing in an image in units of pixels. Image segmentation is widely used in many industrial fields. With the recent development of deep learning, many deep-learning-based image segmentation algorithms have been proposed. Such deep-learning-based image segmentation methods greatly exceed the performance of traditional feature-based image segmentation algorithms in terms of quantitative evaluation such as F1-score. However, deep-learning-based methods require datasets for network training, which has the disadvantage of consuming a lot of time and resources to generate the data. Therefore, the quality and quantity of data collection and annotation are determined by the available resources, which leads to imbalanced data problems. In the imbalanced data environment, the network is trained biasedly according to the size of the class, causing problems such as zero convergence problems. In the case of using weakly data to overcome this drawback, the image segmentation network shows the result of including the noise tendency due to noise contained in the weakly data. To overcome these shortcomings, in this dissertation, a deep-learning-based network learning method using a ternary classifier is proposed that we call IDIS (Imbalanced Dataset Image Segmentation method). The IDIS improves the performance of the image segmentation network by adjusting the relationship among the datasets (background, accurate annotation, and weakly dataset) using a ternary classifier. The ternary classifier in IDIS discriminates the inequality relationship among the datasets by shape and thickness of input, and the IDIS utilizes it in adversarial training step to prevent the adverse effects of training caused by the imbalanced dataset. To evaluate the performance of the proposed method we conducted an experiment on a dataset for detecting cracks on the inner wall of a tunnel where an extreme data imbalance occurs. Experimental results have shown that the proposed IDIS significantly reduces the false positive rate, and improves the performance (as measured by the F1-score) by 0.28, compared with previous methods. In addition, by applying the proposed method to the pavement crack detection problem with imbalanced datasets, the effectiveness of the proposed method in other domains is also demonstrated.