Recent advances on loss functions in deep learning for computer vision

The loss function, also known as cost function, is used for training a neural network or other machine learning models. Over the past decade, researchers have designed many loss functions for machine learning, such as mean squared error and mean absolute error. However, in deep learning, neurons of the last layer are usually activated by a sigmoid or softmax function. Thus, training with traditional losses would cause lower efficiency and accuracy. Recently, designing loss functions for deep learning methods has become one of the most challenging problems. This paper provides a comprehensive review of the recent progress and frontiers about loss functions in deep learning, especially for computer vision tasks. Specifically, we discuss the loss functions in three main computer vision tasks, i.e., object detection, face recognition, and image segmentation. Scholars have proposed several novel loss functions to cope with the specific problems such as imbalanced data, uncertain distribution of the predicted bounding boxes, varied overlapped mode between two bounding boxes and over-fitting. The survey details the source, derivation, and properties of each loss function. Furthermore, we also provide some advanced challenges about robust losses, generative adversarial networks, noise-tolerant losses, and semantic data augmentation. Finally, we deliver a summary and some promising future research directions.