A Comparative Analysis of Loss Functions for Handling Foreground-Background Imbalance in Image Segmentation

Foreground-Background (F-B) imbalance problem has emerged as a fundamental challenge to building accurate image segmentation models in computer vision. F-B imbalance problem occurs due to a disproportionate ratio of observations of foreground and background samples. It degrades the model performance of image segmentation applications. Several loss functions are proposed to improve the classification performance of models on the traditional imbalanced data. But there is no sufficient information on whether these loss functions can improve the accuracy performance of the image segmentation models in the presence of F-B imbalance problem. In this paper, we perform a comparative analysis between four loss functions, namely Focal loss, Dice loss, Tversky and Mixed Focal loss to handle the F-B imbalance problem on datasets from various domains. We embed each loss function in the deep learning model using U-Net architecture and experimentally perform the evaluation in terms of accuracy, precision, F1-score and recall metrics. Then, we present a comparative discussion to the researchers to identify the appropriate loss function that achieves better detection accuracy in different domains.