Bayesian losses for homoscedastic aleatoric uncertainty modeling in pollen image detection

The paper investigates the homoscedastic aleatoric uncertainty modeling for the detection of pollen in images. The new uncertainty modeling loss functions are presented, which are based on the focal and smooth L1 losses. The focal and smooth L1 losses proved their efficiency for the problem of image detection, however, they do not allow modeling the aleatoric uncertainty, while the proposed functions do, leading to more accurate solutions. The functions are based on Bayesian inference and allow for effortless use in existing neural network detectors based on the RetinaNet architecture. The advantages of the loss functions are described on the problem of pollen detection in images. The new loss functions increased the accuracy of pollen image detection, namely localization and classification, on average by 2.76 %, which is crucial for the pollen recognition in general. This helps to automate the process of determining allergenic pollen in the air and reduce the time to inform patients with pollinosis to prevent allergy symptoms. The obtained result shows that the modeling of homoscedastic aleatoric uncertainty for neural networks allows separating the noise from the data, increasing the accuracy of the proposed solutions. The developed functions can be applied to train neural network detectors on any other image datasets.