Dynamic adjustment of hyperparameters for anchor-based detection of objects with large image size differences

摘要

This paper investigates dynamic training for anchor-based detection of objects with large image size dif-ferences. We define different hyper-parameters for training according to different image sizes of objects. By monitoring and collecting the anchor parameters during training, the values of the hyper-parameters for labeling and regression of samples are adjusted dynamically. The dynamic labeling and dynamic re-gression of samples are combined into one training process. Better balance of positive and negative sam-ples is achieved. The quantity and quality of candidate anchors are ensured. It is avoided that candidate anchors miss their objects caused by inappropriate hyper-parameters. The ability to detect objects with different image sizes is improved, especially for small objects (objects with small image sizes). The ex-perimental results show the effectiveness of the proposed strategy for detecting objects with large image size differences (especially small objects) in complex scenes.(c) 2023 Elsevier B.V. All rights reserved.